Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

2102.05095

271

bLVNet (Fan et al., 2019) 65.2 N/A TimeSformer 59.5 74.9 TimeSformer-HR 62.2 78.0 TimeSformer-L 62.4 81.0 Table 7. Video-level accuracy on Something-Something-V2 and Diving-48.Due to an issue with Diving-48 labels used in previously published results, we only compare our method with a reproduced SlowFast 168R101 model. All models are pretained on ImageNet-1K. about 12seconds of a Kinetics video with a single clip. Something-Something-V2 & Diving-48. In Table 7, we also validate our model on SSv2 and Diving-48. Since ImageNet-21K pretraining does not improve accuracy on SSv2 (see Table 3), in this case, we use TimeSformer pretrained on ImageNet-1K. This also allows us to apply the same pretraining to all other models in this comparison, using a ResNet pretrained on ImageNet-1K. Our results suggest that TimeSformer achieves lower accuracy than the best models on this dataset. However, considering that our model uses a completely different design, we take these results as suggesting that TimesSformer is a promising approach even

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

272

models on this dataset. However, considering that our model uses a completely different design, we take these results as suggesting that TimesSformer is a promising approach even for challenging temporally-heavy datasets, such as SSv2. In Table 7, we also present our method on another “temporally-heavy” dataset, Diving-48. Due to a recently discovered issue with a previous version of Diving-48 labels, here, we only compare our method with a reproduced SlowFast168R101 model. Our results show that TimeSformer outperforms SlowFast by a substantial margin. 4.6. Long-Term Video Modeling Lastly, we evaluate TimeSformer on the task of long-term video modeling using HowTo100M (Miech et al., 2019). HowTo100M is an instructional video dataset that contains around 1M instructional Web videos showing humans performing over 23K different tasks, such as cooking, repairing, making arts, etc. The average duration of these videos is around 7minutes, which is orders of magnitude longer than the duration of videos in standard action recognition benchmarks. Each HowTo100M video has a label indicating the task demonstrated in the video (one out of the 23K classes), which can be used for supervised training. Thus, it is a good

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

273

task demonstrated in the video (one out of the 23K classes), which can be used for supervised training. Thus, it is a good benchmark to assess the ability of a model to recognize activities exhibited over very long temporal extents. For this evaluation, we consider only categories that have at least 100 video examples. This gives a subset of HowTo100M corresponding to 120Kvideos spanning 1059 task categories. We randomly partition this collection into 85Ktraining videos and 35Ktesting videos. Is Space-Time Attention All You Need for Video Understanding? Method # Input FramesSingle Clip Coverage# Test ClipsTop-1 Acc SlowFast 8 8.5s 48 48.2 SlowFast 32 34.1s 12 50.8 SlowFast 64 68.3s 6 51.5 SlowFast 96 102.4s 4 51.2 TimeSformer 8 8.5s 48 56.8 TimeSformer 32 34.1s 12 61.2 TimeSformer 64 68.3s 6 62.2 TimeSformer 96 102.4s 4 62.6 Table 8. Long-term task classification on HowTo100M. Given a video spanning several minutes, the goal is to predict the long-term task demonstrated in the video (e.g., cooking breakfast, cleaning

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

274

video spanning several minutes, the goal is to predict the long-term task demonstrated in the video (e.g., cooking breakfast, cleaning house, etc). We evaluate a few variants of SlowFast and TimeSformer on this task. “Single Clip Coverage” denotes the number of seconds spanned by a single clip. “# Test Clips” is the average number of clips needed to cover the entire video during inference. All models in this comparison are pretrained on Kinetics-400. We present our results in Table 8. As our baselines, we use four variants of SlowFast R101, all operating on video clips sampled at a frame rate of 1=32but having varying number of frames: 8;32;64and96. We use the same four configurations for TimeSformer, starting from a ViT pretrained on ImageNet-21K. All models in this comparison are pretrained on Kinetics-400 before finetuning on HowTo100M. During inference, for each method, we sample as many non-overlapping temporal clips as needed to cover the full temporal extent of a video, e.g., if a single clip spans 8:5 seconds, we would sample 48test clips to cover a video of 410seconds. Video-level classification is done by averaging

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

275

seconds, we would sample 48test clips to cover a video of 410seconds. Video-level classification is done by averaging the clip predictions. From the results in Table 8 we first note that, for the same single clip coverage, TimeSformer outperforms the corresponding SlowFast by a large margin of 811%. We also observe that longer-range TimeSformers do better, i.e., our longest-range variant achieves the best video-level classification accuracy. These results suggest that our model is highly suitable for tasks that require long-term video modeling. We also experimented with finetuning TimeSformer directly from a ViT pretrained on ImageNet-1K and ImageNet21K (skipping the Kinetics-400 training). We report that when pretrained only on ImageNet-1K, our model achieves top-1 accuracies of 52:8;58:4;59:2;59:4for8;32;64;96 frame inputs, respectively. When considering ImagNet21K pretraining, TimeSformer produces top-1 accuracies of 56:0;59:2;60:2;62:1for8;32;64;96frame inputs, respectively. These results demonstrate that our model can effectively exploit long-range temporal dependencies regardless

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

276

of the pretraining dataset that we use. 4.7. Additional Ablations Smaller & Larger Transformers. In addition to the “Base” ViT model (Dosovitskiy et al., 2020), we also experimented with the “Large” ViT. We report that this yielded results 1% Figure 7. Visualization of space-time attention from the output token to the input space on Something-Something-V2. Our model learns to focus on the relevant parts in the video in order to perform spatiotemporal reasoning. ViT-50 -40 -30 -20 -10 0 10 20 30 40 50-50-40-30-20-1001020304050 -60 -40 -20 0 20 40-40-30-20-1001020304050 TimeSformer w/ Divided Space-Time Attention-60 -40 -20 0 20 40 60-60-40-200204060 TimeSformer w/ Space Attention Figure 8. Feature visualization with t-SNE (van der Maaten & Hinton, 2008) on Something-Something-V2. Each video is visualized as a point. Videos belonging to the same action category have the same color. The TimeSformer with divided space-time attention

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

277

as a point. Videos belonging to the same action category have the same color. The TimeSformer with divided space-time attention learns semantically more separable features than the TimeSformer with space-only attention or ViT (Dosovitskiy et al., 2020). worse on both Kinetics-400, and Something-Something-V2. Given that our “Base” model already has 121Mparameters, we suspect that the current datasets are not big enough to justify a further increase in model capacity. We also tried the “Small” ViT variant, which produced accuracies about 5%worse than our default “Base” ViT model. Larger Patch Size. We also experimented with a different patch size, i.e., P= 32 . We report that this variant of our model produced results about 3%worse than our default variant using P= 16 . We conjecture that the performance decrease with P= 32 is due to the reduced spatial granularity. We did not train any models with Pvalues lower than 16 as those models have a much higher computational cost. The Order of Space and Time Self-Attention. Our proposed “Divided Space-Time Attention” scheme applies temporal attention and spatial attention one after the other. Here, we investigate whether reversing the order of time-space

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

278

we investigate whether reversing the order of time-space attention (i.e., applying spatial attention first, then temporal) has an impact on our results. We report that applying spatial attention first, followed by temporal attention leads to a 0:5%drop in accuracy on both Kinetics-400, and Something-Something-V2. We also tried a parallel spacetime self-attention. We report that it produces 0:4%lower accuracy compared to our adopted “Divided Space-Time Attention” scheme. Is Space-Time Attention All You Need for Video Understanding? 4.8. Qualitative Results Visualizing Learned Space-Time Attention. In Figure 7, we present space-time attention visualizations obtained by applying TimeSformer on Something-Something-V2 videos. To visualize the learned attention, we use the Attention Rollout scheme presented in (Abnar & Zuidema, 2020). Our results suggest that TimeSformer learns to attend to the relevant regions in the video in order to perform complex spatiotemporal reasoning. For example, we can observe that the model focuses on the configuration of the hand when visible and the object-only when not visible. Visualizing Learned Feature Embeddings. In Figure 8,

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

279

the model focuses on the configuration of the hand when visible and the object-only when not visible. Visualizing Learned Feature Embeddings. In Figure 8, we also visualize the features learned by TimeSformer on Something-Something-V2. The visualization is done using t-SNE (van der Maaten & Hinton, 2008) where each point represents a single video, and different colors depict different action categories. Based on this illustration, we observe that TimeSformer with divided space-time attention learns semantically more separable features than the TimeSformer with space-only attention or ViT (Dosovitskiy et al., 2020). 5. Conclusion In this work, we introduced TimeSformer, a fundamentally different approach to video modeling compared to the established paradigm of convolution-based video networks. We showed that it is possible to design an effective, and scalable video architecture built exclusively on space-time self-attention. Our method (1) is conceptually simple, (2) achieves state-of-the-art results on major action recognition benchmarks, (3) has low training and inference cost, and (4) can be applied to clips of over one minute, thus enabling

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

280

benchmarks, (3) has low training and inference cost, and (4) can be applied to clips of over one minute, thus enabling long-term video modeling. In the future, we plan to extend our method to other video analysis tasks such as action localization, video captioning and question-answering. Appendix A. Implementation Details Our TimeSformer implementation is built using PySlowFast (Fan et al., 2020) and pytorch-image-models (Wightman, 2019) packages. Below, we describe specific implementation details regarding the training and inference procedures of our model. Training. We train our model for 15epochs with an initial learning rate of 0:005, which is divided by 10at epochs 11;and14. During training, we first resize the shorter side of the video to a random value in [256;320]. We then randomly sample a 224224crop from the resized video. For our high-resolution model, TimeSformer-HR, we resize the shorter side of the video to a random value in [448;512], and then randomly sample a 448448crop. We randomlysample clips from the full-length videos with a frame rate of1=32. The batch size is set to 16. We train all our models

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

281

of1=32. The batch size is set to 16. We train all our models using synchronized SGD across 32GPUs. The momentum is set to 0:9, while the weight decay is set to 0:0001 . Unless otherwise noted, in our experiments we use the “Base” ViT model (Dosovitskiy et al., 2020). Temporal and spatial attention layers in each block are initialized with the same weights, which are obtained from the corresponding attention layer in ViT. Inference. As discussed in the main draft, during inference we sample a single temporal clip in the middle of the video. We scale the shorter spatial side of a video to 224pixels (or 448for TimeSformer-HR) and take 3crops of size 224224 (448448for TimeSformer-HR) to cover a larger spatial extent within the clip. The final prediction is obtained by averaging the softmax scores of these 3predictions. Other models in our comparison. To train I3D (Carreira & Zisserman, 2017), and SlowFast (Feichtenhofer et al., 2019b), we use the training protocols that were used in the original papers. For I3D, we initialize it with a 2D ImageNet

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

282

2019b), we use the training protocols that were used in the original papers. For I3D, we initialize it with a 2D ImageNet CNN, and then train it for 118epochs with a base learning rate of 0:01, which is divided by 10at epochs 44and88. We use synchronized SGD across 32GPUs following the linear scaling recipe of Goyal et al. (2017a). We set the momentum to 0:9, and weight decay to 0:0001 . The batch size is set to 64. For the SlowFast model, when initialized from ImageNet weights, we use this same exact training protocol. When training SlowFast from scratch, we use the training protocol described by the authors (Feichtenhofer et al., 2019b). More specifically, in that case, the training is done for 196epochs with a cosine learning rate schedule, and the initial learning rate is set to 0:1. We use a linear warm-up for the first 34epochs starting with a learning rate of0:01. A dropout of 0:5is used before the final classification layer. The momentum is set to 0:9, the weight decay is

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

283

0:0001 , and the batch size is set to 64. Just as before, we adopt the linear scaling recipe (Goyal et al., 2017a). Datasets. Kinetics-400 (Carreira & Zisserman, 2017) consists of 240Ktraining videos and 20Kvalidation videos that span 400human action categories. Kinetics-600 (Carreira et al., 2018) has 392Ktraining videos and 30Kvalidation videos spanning 600action categories. SomethingSomething-V2 (Goyal et al., 2017b) contains 170Ktraining videos and 25Kvalidation videos that span 174action categories. Lastly, Diving-48 (Li et al., 2018) has 16Ktraining videos and 3Ktesting videos spanning 48fine-grained diving categories. For all of these datasets, we use standard classification accuracy as our main performance metric. Is Space-Time Attention All You Need for Video Understanding? References Abnar, S. and Zuidema, W. Quantifying attention flow in transformers, 2020. Ba, L. J., Kiros, J. R., and Hinton, G. E. Layer normalization. CoRR , 2016. Bello, I., Zoph, B., Le, Q., Vaswani, A., and Shlens, J.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

284

CoRR , 2016. Bello, I., Zoph, B., Le, Q., Vaswani, A., and Shlens, J. Attention augmented convolutional networks. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV , 2019. Bertasius, G. and Torresani, L. Classifying, segmenting, and tracking object instances in video with mask propagation. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020. Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-V oss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D. Language models are few-shot learners. 2020.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

285

S., Radford, A., Sutskever, I., and Amodei, D. Language models are few-shot learners. 2020. Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., and Zagoruyko, S. End-to-end object detection with transformers. In European Conference Computer Vision (ECCV) , 2020. Carreira, J. and Zisserman, A. Quo vadis, action recognition? A new model and the kinetics dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , 2017. Carreira, J., Noland, E., Banki-Horvath, A., Hillier, C., and Zisserman, A. A short note about kinetics-600. CoRR , 2018. Chen, M. X., Firat, O., Bapna, A., Johnson, M., Macherey, W., Foster, G., Jones, L., Schuster, M., Shazeer, N., Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Chen,

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

286

Z., Wu, Y ., and Hughes, M. The best of both worlds: Combining recent advances in neural machine translation. InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics . Association for Computational Linguistics, 2018a. Chen, Y ., Kalantidis, Y ., Li, J., Yan, S., and Feng, J. Aˆ2nets: Double attention networks. In Advances in Neural Information Processing Systems 31 , 2018b. Chen, Y ., Fan, H., Xu, B., Yan, Z., Kalantidis, Y ., Rohrbach, M., Yan, S., and Feng, J. Drop an octave: Reducing spatial redundancy in convolutional neural networks with octave convolution. In Proceedings of the IEEE/CVFInternational Conference on Computer Vision (ICCV) , October 2019. Child, R., Gray, S., Radford, A., and Sutskever, I. Generating long sequences with sparse transformers. CoRR , 2019. Cordonnier, J., Loukas, A., and Jaggi, M. On the relationship between self-attention and convolutional layers. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30,

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

287

8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 , 2020. Dai, Z., Yang, Z., Yang, Y ., Carbonell, J., Le, Q., and Salakhutdinov, R. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019. Deng, J., Dong, W., Socher, R., Li, L., Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition , pp. 248–255, 2009. doi: 10.1109/CVPR. 2009.5206848. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 , 2018. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT:

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

288

2018. Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , 2019. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR , 2020. Fan, H., Li, Y ., Xiong, B., Lo, W.-Y ., and Feichtenhofer, C. Pyslowfast. https://github.com/ facebookresearch/slowfast , 2020. Fan, Q., Chen, C.-F. R., Kuehne, H., Pistoia, M., and Cox, D. More is less: Learning efficient video representations by big-little network and depthwise temporal aggregation.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

289

D. More is less: Learning efficient video representations by big-little network and depthwise temporal aggregation. InAdvances in Neural Information Processing Systems , volume 32, 2019. Feichtenhofer, C. X3d: Expanding architectures for efficient video recognition. CVPR , pp. 200–210, 2020. Feichtenhofer, C., Fan, H., Malik, J., and He, K. Slowfast networks for video recognition. In Proceedings of the Is Space-Time Attention All You Need for Video Understanding? IEEE/CVF International Conference on Computer Vision (ICCV) , 2019a. Feichtenhofer, C., Fan, H., Malik, J., and He, K. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV , 2019b. Gavrilyuk, K., Sanford, R., Javan, M., and Snoek, C. G. M. Actor-transformers for group activity recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 2020. Girdhar, R., Carreira, J., Doersch, C., and Zisserman, A. Video action transformer network. In IEEE Conference on

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

290

Girdhar, R., Carreira, J., Doersch, C., and Zisserman, A. Video action transformer network. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR , 2019. Goyal, P., Dollár, P., Girshick, R., Noordhuis, P., Wesolowski, L., Kyrola, A., Tulloch, A., Jia, Y ., and He, K. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677 , 2017a. Goyal, R., Kahou, S. E., Michalski, V ., Materzynska, J., Westphal, S., Kim, H., Haenel, V ., Fründ, I., Yianilos, P., Mueller-Freitag, M., Hoppe, F., Thurau, C., Bax, I., and Memisevic, R. The "something something" video database for learning and evaluating visual common sense. CoRR , 2017b. Ho, J., Kalchbrenner, N., Weissenborn, D., and Salimans, T. Axial attention in multidimensional transformers. CoRR , 2019.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

291

Axial attention in multidimensional transformers. CoRR , 2019. Hu, H., Gu, J., Zhang, Z., Dai, J., and Wei, Y . Relation networks for object detection. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR , 2018. Huang, Z., Wang, X., Huang, L., Huang, C., Wei, Y ., and Liu, W. Ccnet: Criss-cross attention for semantic segmentation. 2019. Jiang, B., Wang, M., Gan, W., Wu, W., and Yan, J. Stm: Spatiotemporal and motion encoding for action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2019. Kwon, H., Kim, M., Kwak, S., and Cho, M. Motionsqueeze: Neural motion feature learning for video understanding. InECCV , 2020. Le, H., Sahoo, D., Chen, N., and Hoi, S. Multimodal transformer networks for end-to-end video-grounded dialogue systems. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

292

systems. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , 2019. Li, L., Chen, Y .-C., Cheng, Y ., Gan, Z., Yu, L., and Liu, J. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200 , 2020a.Li, Y ., Li, Y ., and Vasconcelos, N. Resound: Towards action recognition without representation bias. In The European Conference on Computer Vision (ECCV) , September 2018. Li, Y ., Ji, B., Shi, X., Zhang, J., Kang, B., and Wang, L. Tea: Temporal excitation and aggregation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020b. Lin, J., Gan, C., and Han, S. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE International Conference on Computer Vision , 2019. Miech, A., Zhukov, D., Alayrac, J.-B., Tapaswi, M., Laptev,

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

293

International Conference on Computer Vision , 2019. Miech, A., Zhukov, D., Alayrac, J.-B., Tapaswi, M., Laptev, I., and Sivic, J. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV , 2019. Ott, M., Edunov, S., Grangier, D., and Auli, M. Scaling neural machine translation. In Proceedings of the Third Conference on Machine Translation: Research Papers , 2018. Parmar, N., Vaswani, A., Uszkoreit, J., Kaiser, L., Shazeer, N., Ku, A., and Tran, D. Image transformer. In Dy, J. G. and Krause, A. (eds.), Proceedings of the 35th International Conference on Machine Learning, ICML , 2018. Qiu, Z., Yao, T., Ngo, C.-W., Tian, X., and Mei, T. Learning spatio-temporal representation with local and global diffusion. In CVPR , 2019. Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. Improving language understanding by generative pretraining. 2018.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

294

Radford, A., Narasimhan, K., Salimans, T., and Sutskever, I. Improving language understanding by generative pretraining. 2018. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. Language models are unsupervised multitask learners. 2019. Ramachandran, P., Parmar, N., Vaswani, A., Bello, I., Levskaya, A., and Shlens, J. Stand-alone self-attention in vision models. In Advances in Neural Information Processing Systems , pp. 68–80, 2019. Sevilla-Lara, L., Zha, S., Yan, Z., Goswami, V ., Feiszli, M., and Torresani, L. Only time can tell: Discovering temporal data for temporal modeling. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pp. 535–544, January 2021. Simonyan, K. and Zisserman, A. Very deep convolutional networks for large-scale image recognition. In ICLR , 2015. Is Space-Time Attention All You Need for Video Understanding?

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

295

networks for large-scale image recognition. In ICLR , 2015. Is Space-Time Attention All You Need for Video Understanding? Stroud, J., Ross, D., Sun, C., Deng, J., and Sukthankar, R. D3d: Distilled 3d networks for video action recognition. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , March 2020. Sun, C., Myers, A., V ondrick, C., Murphy, K., and Schmid, C. Videobert: A joint model for video and language representation learning, 2019. Szegedy, C., Liu, W., Jia, Y ., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V ., and Rabinovich, A. Going deeper with convolutions. In Computer Vision and Pattern Recognition (CVPR) , 2015. Teed, Z. and Deng, J. RAFT: recurrent all-pairs field transforms for optical flow. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II , 2020.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

296

- 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II , 2020. Tran, D., Wang, H., Torresani, L., Ray, J., LeCun, Y ., and Paluri, M. A closer look at spatiotemporal convolutions for action recognition. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018 , 2018. Tran, D., Wang, H., Feiszli, M., and Torresani, L. Video classification with channel-separated convolutional networks. ICCV , pp. 5551–5560, 2019. van der Maaten, L. and Hinton, G. Visualizing data using t-SNE. Journal of Machine Learning Research , 9: 2579–2605, 2008. URL http://www.jmlr.org/ papers/v9/vandermaaten08a.html . Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Advances in Neural Information Processing Systems , 2017a.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

297

Processing Systems , 2017a. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. Attention is all you need. In Advances in Neural Information Processing Systems 30 . 2017b. Wang, H., Tran, D., Torresani, L., and Feiszli, M. Video modeling with correlation networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020a. Wang, H., Zhu, Y ., Green, B., Adam, H., Yuille, A. L., and Chen, L. Axial-deeplab: Stand-alone axial-attention for panoptic segmentation. In Computer Vision - ECCV 2020 - 16th European Conference , 2020b. Wang, X., Girshick, R., Gupta, A., and He, K. Non-local neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018a.Wang, X., Girshick, R. B., Gupta, A., and He, K. Non-local

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

298

June 2018a.Wang, X., Girshick, R. B., Gupta, A., and He, K. Non-local neural networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , 2018b. Wang, X., Xiong, X., Neumann, M., Piergiovanni, A. J., Ryoo, M. S., Angelova, A., Kitani, K. M., and Hua, W. Attentionnas: Spatiotemporal attention cell search for video classification. In Computer Vision - ECCV 2020 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII , 2020c. Weissenborn, D., Täckström, O., and Uszkoreit, J. Scaling autoregressive video models. In 8th International Conference on Learning Representations, ICLR , 2020. Wightman, R. Pytorch image models. https://github. com/rwightman/pytorch-image-models , 2019. Xie, S., Sun, C., Huang, J., Tu, Z., and Murphy, K.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

299

com/rwightman/pytorch-image-models , 2019. Xie, S., Sun, C., Huang, J., Tu, Z., and Murphy, K. Rethinking spatiotemporal feature learning: Speedaccuracy trade-offs in video classification. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part XV , pp. 318–335, 2018. doi: 10.1007/ 978-3-030-01267-0\_19. URL https://doi.org/ 10.1007/978-3-030-01267-0_19 . Yang, Z., Garcia, N., Chu, C., Otani, M., Nakashima, Y ., and Takemura, H. Bert representations for video question answering. In The IEEE Winter Conference on Applications of Computer Vision , 2020. Zhao, H., Jia, J., and Koltun, V . Exploring self-attention for image recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR , 2020. Zhou, L., Zhou, Y ., Corso, J. J., Socher, R., and Xiong, C.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2102.05095

300

Zhou, L., Zhou, Y ., Corso, J. J., Socher, R., and Xiong, C. End-to-end dense video captioning with masked transformer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2018.

2102.05095

Is Space-Time Attention All You Need for Video Understanding?

We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically new design, TimeSformer achieves state-of-the-art results on several action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Finally, compared to 3D convolutional networks, our model is faster to train, it can achieve dramatically higher test efficiency (at a small drop in accuracy), and it can also be applied to much longer video clips (over one minute long). Code and models are available at: https://github.com/facebookresearch/TimeSformer.

http://arxiv.org/pdf/2102.05095

Accepted to ICML 2021

cs.CV

20210209

20210609

2103.15662

0

Unified Graph Structured Models for Video Understanding Anurag Arnab Chen Sun Cordelia Schmid Google Research faarnab, chensun, cordelias [email protected] Abstract Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model’s performance. We demonstrate our method on two different tasks requiring relational reasoning in videos – spatio-temporal action detection on AVA and UCF10124, and video scene graph classification on the recent Action Genome dataset – and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene. 1. Introduction Deep learning has enabled rapid advances in many image understanding tasks such as image classification [20], object detection [46] and semantic segmentation [7]. However, progress on recent video understanding datasets such

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

1

object detection [46] and semantic segmentation [7]. However, progress on recent video understanding datasets such as A V A [17] and Charades [49] has lagged behind in comparison. Further progress in the video understanding tasks posed by these datasets would facilitate applications in autonomous vehicles, health monitoring and automated media analysis and production among others. A reason why video understanding is so challenging is because, as shown in Fig. 1, it requires understanding the interactions between actors, objects and other context in the scene. Furthermore, these interactions are not always observable from a single frame, and thus require reasoning over long temporal intervals. This is illustrated in Fig. 1, where understanding the actions of the person in the centreframe is not possible from the target keyframe alone. In order to know that the woman is “listening”, we need to con-sider the man who is speaking but no longer in the scene. And to correctly infer that the woman is “driving” the car, rather than “riding” like the man, we must take into account that she is later holding the steering wheel. Video is a significantly higher-dimensional signal than single images, due to its additional temporal axis, and so we believe learning these unlabelled interactions directly from current datasets with large convolutional networks is not

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

2

single images, due to its additional temporal axis, and so we believe learning these unlabelled interactions directly from current datasets with large convolutional networks is not feasible. In this paper, we propose a structured graph neural network to explicitly model these spatio-temporal interactions. We model actors and objects (explicitly with bounding boxes when we have supervision, and implicitly otherwise) as nodes in our spatio-temporal graph and perform message passing inference to directly model their relations. Although a wide range of graph-structured models have been proposed for action recognition, we note that there has been no unifying formulation for these models. As such, some works only model spatial relations between actors and objects [14, 54], but not how these interactions evolve over time. Other approaches model long-range temporal interactions [61], but do not capture spatial relations and are not trained end-to-end. And whilst some methods do model spatio-temporal interactions of objects [3, 60], their explicit representations of objects need additional supervision, and are not evaluated on spatio-temporal localisation tasks which requires detailed understanding and is necessary for analysing untrimmed videos. Our graph network formulation based on the messagepassing neural network [13] abstraction, allows us to explicitly model interactions between actors, objects and

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

3

Our graph network formulation based on the messagepassing neural network [13] abstraction, allows us to explicitly model interactions between actors, objects and the scene, and how these interactions evolve over time. Our flexible model allows us to use explicit object representations from a pretrained Region Proposal Network (RPN) [46], and/or implicitly from convolutional feature maps without additional supervision. Moreover, our general formulation allows us to interpret previous work [14, 54, 60, 61, 71] as special cases, and thus understand how different design choices in object representation, graph connectivity and message passing functions affect the model’s performance. We demonstrate our versatile model on two different tasks: spatio-temporal action detection on A V A [17]arXiv:2103.15662v1 [cs.CV] 29 Mar 2021 TimeTarget keyframeLabels: listen, drive, sit…………Figure 1: Understanding videos requires reasoning over long-term spatio-temporal interactions between actors, objects and the environment. The actions of the woman in the centre frame are ambiguous given the nearby frames which typical 3D CNN architectures consider. However, by considering her interactions with the man from nearby frames, we know she is “listening” to him. And the fact that she is later

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

4

However, by considering her interactions with the man from nearby frames, we know she is “listening” to him. And the fact that she is later holding the steering wheel indicates she is “driving” in contrast to the man who is “riding”. In this paper, we propose a spatio-temporal graph on which we perform message passing to explicitly model these spatio-temporal interactions. Example from the A V A dataset [17]. and UCF101-24 [53], and video scene graph prediction on the recent Action Genome [23] dataset. Both of these tasks require modelling the spatio-temporal interactions between the actors and/or objects in the scene, and we show consistent improvements from using each component of our model and achieve state-of-the-art results on each dataset. Furthermore, we observe that the largest improvements in A V A are indeed achieved on action classes involving human-to-human and human-to-object interactions, and visualisations of our network show that it is focusing on scene context that is intuitively relevant to its action classification. 2. Related Work Modelling contextual relationships has a long history in scene understanding. Relevant examples of early works in this area included modelling interactions between humans and objects [18, 68], different objects [45] and relationships

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

5

scene understanding. Relevant examples of early works in this area included modelling interactions between humans and objects [18, 68], different objects [45] and relationships between human actions and scene context [41, 16]. Furthermore, it has also been shown that human vision is reliant on context too [43]. In this paper, we consider video understanding tasks, specifically spatio-temporal action recognition and video scene graph parsing, which involve reasoning about interactions between actors, objects and their environment in both space and time. Early work in action recognition used hand-crafted features to encode motion information [32, 58]. Advances in deep learning first saw repurposing of 2D image convolutional neural networks (CNNs) for video as “two stream” networks [27, 50] followed by spatio-temporal 3D CNNs [6, 12, 55, 72]. However, these architectures focus on extracting coarse, video-level features and are not suitable for learning the fine-grained relations depicted in Fig. 1. Consequently, whilst initial approaches to spatiotemporal action detection involved extending 2D object detectors [38, 46] temporally [26, 51, 1, 65], current leading methods [14, 61, 71] on the A V A dataset [17] all explicitly

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

6

methods [14, 61, 71] on the A V A dataset [17] all explicitly model relationships, with approaches which we show can be interpreted as variants of graph neural networks. Graph neural networks (GNNs) explicitly model relations between entities by modelling them as nodes in adirected or undirected graph [5, 48, 29], which interact via a neighbourhood defined for each node. The selfattention [56] and Non-local [59] operators can also be thought of as GNNs [5, 13] where each element in a feature map is a node, and all nodes are fully-connected to each other. Such attention-based models have excelled at a number of natural language processing and computer vision tasks and have inspired many follow-up methods [21, 8, 9, 62, 70]. Many structured models have also recently been employed in video understanding. However, there has previously been no coherent framework to unify these approaches. Consequently, some works only model spatial relations between actors and objects, but not how these evolve over time [54, 14]. And whilst LFB [61] models long range temporal interactions, it does not capture spatial relationships within a keyframe. Moreover, in order to model longrange interactions, [61] uses precomputed features and is thus not trained end-to-end. Our proposed method, which is

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

7

thus not trained end-to-end. Our proposed method, which is based on the message passing neural network (MPNN) [13] framework, coherently models both spatial and temporal interactions. After describing our model in Sec. 3, we show how previous structured models for action detection [54, 14, 60, 61, 71] can be thought of as special cases of our model. Furthermore, our flexible model can reason about objects both when we have explicit supervision for them, and when we do not. Prior work in video understanding, which have proposed less generic models, have either assumed the case of explicit object supervision [3, 60] or not [54, 14, 71], but have not considered the scenario when both options are available. Furthermore, our general formulation allows us to ablate graph modelling design choices, such as the object representation, message passing functions and temporal context, in a manner not possible with these previous, more specific approaches. Moreover, we note that some approaches which have employed spatiotemporal graphs in video have only considered frame-toframe interactions [22, 60, 3, 40, 42, 44], rather than longrange relations, and have not evaluated on spatio-temporal localisation which requires more detailed understanding 3D CNN sit, answer phone, carry/hold object

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

8

localisation which requires more detailed understanding 3D CNN sit, answer phone, carry/hold object sit, touch (an object), watch (a person), talk (a person) sit, touch (an object), watch (a person), listen (a person) Readout Function, Spatio temporal graph construction and message passing Foreground node Implicit context Explicit contextFigure 2: Overview of our method: We construct a spatio-temporal graph, and perform message-passing inference on it, to model interactions between actors, objects and their environment. Foreground nodes (circles) have readout functions associated with them for the task of interest ( i.e. for action recognition, the nodes represent person bounding boxes which are classified into actions). Context nodes (squares) model additional information, and can either be implicit, as cells of the original feature map, or explicit by ROI-pooling external region proposals (RPN [46] not shown for clarity). The initial state of each node is a spatio-temporal feature vector extracted from a 3D CNN. and is essential for analysing untrimmed videos. We also note that scene graph parsing [25, 30] is another task that evaluates a model’s ability to reason about

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

9

and is essential for analysing untrimmed videos. We also note that scene graph parsing [25, 30] is another task that evaluates a model’s ability to reason about the interactions between different objects by representing objects as nodes and relationships as edges in a graph. Although the task was originally posed for single images [30], the recent Action Genome [23] dataset extends this task to video by adding annotations to Charades [49]. While GNN-based approaches have also been used in scene graph parsing [10, 33, 64, 66] for single images, to our knowledge, they have not been applied to model temporal relations in video. Moreover, the same model has not also been demonstrated on spatio-temporal action recognition like our method. 3. Proposed Approach Our model aims to build a structured representation of a video by representing it as a graph of actors, objects and contextual elements in the scene, as shown in Fig. 2. This structured representation is then used to perform tasks which require understanding the interactions between the elements in the graph, such as action recognition and scene graph prediction. Note that we do not assume we have annotations for relevant scene context. Our approach is based on Message Passing Neural Networks (MPNN) [13], as it is a flexible framework that generalises many previous graph neural network algorithms [4,

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

10

Our approach is based on Message Passing Neural Networks (MPNN) [13], as it is a flexible framework that generalises many previous graph neural network algorithms [4, 5, 29, 57] We review this approach, and describe how we adapt it for video understanding in Sec. 3.1, before detailing our model in Sec. 3.2 through 3.5. Finally, we dis-cuss how previous structured models for video understanding [14, 54, 60, 61, 71] can be regarded as specific instantiations of our model in Sec. 3.6. 3.1. Message Passing Neural Networks (MPNNs) MPNNs operate on a directed or undirected graph, G consisting of nodes, v2V, and a neighbourhood for each node,Nv, that defines the graph’s connectivity. For video models we distinguish between the spatial, Sv, and temporal,Tv, neighbourhoods for node v(Nv=Sv[Tv). Each node,v, is associated with a latent state, hv. Inference in this model consists of a message passing phase, and a final readout phase. In the message passing phase, messages for each node, mv, are first computed by applying spatial

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

11

for each node, mv, are first computed by applying spatial and temporal message passing functions, MsandMtrespectively, to all nodes in its neighbourhood as described in Eq. (1). An update function, U, then aggregates the received messages to update the latent state, hv, mi+1 v=X w2SvMs(hi v;hi w;i s) +X u2TvMt(hi v;hi u;i t)(1) hi+1 v=U(mi+1 v;hi v); (2) wheredenotes learnable function parameters. Intuitively, the state of a node, hv, is updated by aggregating the messages passed from its neighbours. Finally, after I1iterations of message passing, a readout function, R, uses the updated node features for the classification task of interest y=R(fhi vgjv2G): (3) As illustrated in Fig. 2, our graph consists of a set of “Foreground” nodes, F=ff1;f2;:::;fngand “Context”

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

12

“Foreground” nodes, F=ff1;f2;:::;fngand “Context” nodes,C=fc1;c2;:::;cmgwherenandmvary for each video. The “Foreground” nodes, F, have readout and loss functions associated with them, and correspond to the objects that will subsequently be classified. For the task of action recognition,Fcorresponds to bounding boxes of each actor in the keyframe. Whilst for scene graph prediction, F refers to bounding boxes for all potential objects of interest. The “Context” nodes, C, capture additional information extracted from the scene for relational reasoning. The representations of these nodes, and their spatial connectivity, are detailed next. 3.2. Spatial Model The spatial connections in our graph model relationships between actors, objects and scene context in the same frame. For example, recognising an action such as “give object to a person” in the A V A dataset [17] requires understanding both people involved in the action and also the object being transferred. To model such interactions, we first represent the foreground nodes of our model, F, by extracting convolutional features from the last layer of the network, X2

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

13

To model such interactions, we first represent the foreground nodes of our model, F, by extracting convolutional features from the last layer of the network, X2 Rthwc, and using ROI-Align [19] followed by spatiotemporal pooling and a linear projection to obtain fi2Rd. We model scene context by considering the features from each spatial position in the feature map, X, as a contextual node,ci, in our graph. Note that these features are projected toRd. A similar representation was used by [54], which we refer to as an implicit object model as it enables the network to encode information about the scene and relevant objects without any extra supervision. This approach is also known as “grid features” in visual question answering [4, 24]. It is also possible to augment our set of contextual nodes with an explicit object representation by computing classagnostic object proposals with a Region Proposal Network (RPN) [46]. We use an RPN pretrained on the OpenImages dataset [31] and obtain a d-dimensional feature from each proposal by using ROI-Align and a linear projection as for the foreground nodes. A similar idea of using explicit object representations in video has also been employed by [3] and

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

14

the foreground nodes. A similar idea of using explicit object representations in video has also been employed by [3] and [60], though not for spatio-temporal action recognition. Messages are then computed and passed to foreground nodes from both foreground and contextual nodes in the graph. We only update the foreground nodes during message passing, as these are the nodes which are subsequently classified in the final Readout phase. Concretely, the spatial neighbourhood for each foreground node in our graph is Sv=F[C8v2F: (4) ContextForegroundFrames ⚫ ⚫ ⚫𝜏"𝜏#Figure 3: Illustration of temporal connectivity. Foreground (blue) and context (green) nodes have initial states computed from spatiotemporal features over tframes (shaded in yellow), centred on a keyframe (denoted by a black circle). scontrols the distance between selected keyframes; cis the total number of keyframes considered. Here, c= 3,s= 7frames andt= 5 frames. 3.3. Temporal Model We also include temporal connections in our graph to model long-range interactions between actors and objects.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

15

3.3. Temporal Model We also include temporal connections in our graph to model long-range interactions between actors and objects. As shown in Fig. 1, understanding actions often requires reasoning about actors who are no longer visible in the current frame, thus requiring large temporal contexts. We model these temporal interactions by connecting foreground nodes in keyframe twith all other foreground nodes in neighbouring keyframes t02T . Concretely, we define the temporal context, c, and temporal stride, s, hyperparameters. As shown in Fig. 3, cis the total number of number of keyframes in the video which we consider in our temporal graph. sis the sampling rate at which we select keyframes, as s1allows us to consider a wider temporal interval in a more computationally efficient manner. This is necessary to train the entire model end-to-end. Moreover, as each foreground feature node in the graph, fi, is itself a spatio-temporal feature computed over a period of tby a 3D CNN, selecting adjacent keyframes (and effectively settings= 1 keyframe) could result in redundant information being captured by the temporal connections in the

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

16

a 3D CNN, selecting adjacent keyframes (and effectively settings= 1 keyframe) could result in redundant information being captured by the temporal connections in the graph. Note that our definition of keyframe follows common datasets [17, 23] as a frame we aim to classify given surrounding temporal context. More formally, we can describe the neighbourhood of each foreground node, v, for temporal message passing as Tv=bc=2c[ dt=c=2eFt
s: (5) Here, we use the superscript to denote the temporal index, and without loss of generality consider t= 0 to be the centre keyframe, meaning that negative time indices correspond to frames in the past. We set cto be an odd, positive integer to employ an equal-sized temporal window on either side of the centre keyframe. We first perform spatial message passing, before passing messages temporally. This allows information from the context nodes in frame t0,ct0 j, to efficiently propagate to a foreground node ft i(wheret6=t0) viaft0 j, as the foreground nodes are fully-connected temporally. The following section now describes the messages that are passed along the graph described above.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

17

j, as the foreground nodes are fully-connected temporally. The following section now describes the messages that are passed along the graph described above. 3.4. Message Passing Functions We first observe that the Non-local operator or selfattention [56, 59] can be considered as a message passing function in a fully-connected graph, where each node, hv, is an element in the input feature map, H, and the neighbourhood comprises of all other feature elements. And since self-attention is employed with a residual connection [59, 56] and layer normalisation [2, 56], the update function of Non-local when viewed as an MPNN is U=LN(hv+Self-Attention (H)): (6) Similar analysis [5, 13] has shown that Graph Attention Networks (GAT) [57], Relational Networks [47] and many other graph neural networks [48, 29, 35, 4] can also be interpreted as MPNNs. In this paper, we consider Non-local [59] and Graph Attention (GAT) [57] as message passing functions within the update function, Eq. (6). Non-local We modify Non-local [59] to pass messages from all Foreground and Context nodes to only Foreground

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

18

Non-local We modify Non-local [59] to pass messages from all Foreground and Context nodes to only Foreground nodes (Eq. 4), as these are the nodes which are used for the final classification in the Readout phase, M=SoftmaxQK> p d V (7) Q=AWqK= [AjjC]WkV= [AjjC]Wv: Here,A2RndandC2Rmdare matrices where each row is a foreground and context feature node respectively, Mis theRndmatrix of messages received by each node in A,[AjjC]denotes the concatenation of these matrices and Wq,WkandWvare learnable ddprojection matrices. Graph attention (GAT) The graph attention [57] message for a node vis computed as mv=0 @X j2NvijWahj1 A (8) ij=Softmax  w> b[hijjhj]

(9) whereis a ReLU non-linearity and Waandwbare a learnable matrix and vector respectively.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

19

b[hijjhj]

(9) whereis a ReLU non-linearity and Waandwbare a learnable matrix and vector respectively. Parallel messages It is also possible to compute multiple incoming messages in parallel for a node, hv. When usingmultiple Non-local or Graph Attention functions, this corresponds to multi-headed attention [56, 57]. By viewing Nonlocal and GAT as message passing functions, we can also aggregate messages from a combination of these two. In these cases, we aggregate the messages using an attentionweighted convex combination, as performed in Eq. (9). 3.5. Readout Function AfterI1iterations of message passing, a readout function is applied on the Foreground nodes to obtain the final predictions. For action detection, the readout function is a linear classifier operating on each element of F, where each foreground node, fi, corresponds to the features of an actor in the keyframe. For scene graph prediction, the readout function consists of two classifiers: The first linear classifier predicts the object-class label of each foreground node. The second is a function of each pair of foreground nodes, and predicts the relationship label between them. 3.6. Discussion

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

20

node. The second is a function of each pair of foreground nodes, and predicts the relationship label between them. 3.6. Discussion We note that many previous structured models for video understanding can be considered as special cases of our proposed MPNN framework: Girdhar et al . [14] only consider a spatial model, i.e. N=S, using implicit objects and Non-local [60] for message passing. ACRN [54] has the same graph structure, but uses Relational Networks [47] for message passing instead. LFB [61], in contrast, considers only a temporal model, i.e.N=T, using Non-local as the message passing function in a graph that is fully-connected in time. However, [61] do not consider a spatial model to capture interactions between actors in the keyframe. Zhang et al. [71], on the other hand, model both temporal and spatial connections. However, they effectively model three separate graphs: the first models actors in short, 3-second tubelets using GCN [29] for message passing. The other two graphs model actor-actor and actor-object relations using a message passing method similar to GAT [57], but using a hand-defined weighting function rather than a learned

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

21

two graphs model actor-actor and actor-object relations using a message passing method similar to GAT [57], but using a hand-defined weighting function rather than a learned one as in GAT (Eq. 9). Wang et al. [60] also model a spatiotemporal graph, using GCN for message passing, and an explicit object representation in the spatial model. However, their temporal connections are only among adjacent frames, which does not allow information propagation between all frames in long sequences, in contrast to our model which is fully-connected temporally. With our unified framework, we study the effect of graph modelling design choices and show how we outperform previous work next in Sec. 4. 4. Experiments We evaluate our method’s ability to model spatiotemporal interactions with experiments on spatio-temporal action detection and video scene graph classification. Table 1: Ablation study on Action Genome using a 3D ResNet 50 backbone. We report the effect of (a) different message passing functions, (b) temporal connections in the graph ( c= 1;s= 1corresponds to only spatial message passing) and (c) iterations of message passing. (a) Spatial message passing functions SGCls PredCls

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

22

(a) Spatial message passing functions SGCls PredCls R@20 R@50 R@10 R@20 Baseline 48.9 51.3 78.7 93.8 Non-local in backbone 49.1 51.4 78.8 93.9 Non-local 50.4 52.6 79.3 94.2 GAT 51.1 53.2 79.7 94.4 GAT + Non-local 51.3 53.4 79.4 94.3(b) Temporal message passing Temporal parameters SGCls PredCls cs R@20 R@50 R@10 R@20 1 1 51.1 53.2 79.7 94.4 3 2 52.9 55.0 79.4 94.2 3 5 53.3 55.5 79.4 94.2 3 7 53.5 55.7 79.3 94.2 5 2 53.4 55.5 79.4 94.2 5 5 53.8 56.0 79.3 94.2 5 7 53.6 55.8 79.3 94.2(c) Message passing iterations Iterations R@20 1 51.1 2 51.6 3 51.6 5 51.8 Spatio-temporal action recognition We evaluate on

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

23

Iterations R@20 1 51.1 2 51.6 3 51.6 5 51.8 Spatio-temporal action recognition We evaluate on A V A [17], the largest dataset for this task consisting of 15-minute video clips obtained from movies, and UCF10124 [53], the previous standard benchmark for this task. A V A is labelled with atomic actions, where one person is typically performing multiple actions simultaneously, whilst actors perform only a single high-level action in UCF101-24. We follow standard protocol and evaluate using the Frame AP at an IoU threshold of 0.5 on both datasets. For A V A, we use v2.2 annotations for ablations, and either v2.1 or v2.2 annotations for fair comparisons with prior work. Video scene graph prediction We evaluate on the recent Action Genome dataset [23], which adds scene graph annotations to Charades [49], on two scene graph tasks: scene graph classification (SGCls) and predicate classification (PredCls). Both tasks are evaluated using the standard Recall@K metric [39, 23] (R@K) which measures the fraction of ground truth relationship triplets (subject-predicateobject) that appear in the top K scoring predicted triplets. In

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

24

SGCls, ground truth bounding box co-ordinates are given, and the aim is to predict their object classes, as well as relationship labels between pairs of objects. PredCls is simpler, as both bounding box co-ordinates and object classes are given, and only the relationship label must be predicted. Implementation details We use the public implementation of SlowFast [11] as our baseline with the 3D ResNet 50 or ResNet 101 backbones [20], as it is the current stateof-the-art. The network is similar to Fast-RCNN [15] as it uses external region proposals to extract features from the last feature map of res 5using ROI-Align [19]. These features are then spatio-temporally pooled and classified. In our graph model, we initialise the internal states, hi, of our Foreground nodes using these ROI-Aligned res 5features. For our action detection experiments, the Foreground nodes in our graph correspond to bounding boxes of actors, and we use the same person detector as [11, 61] for our actor region proposals. For scene graph experiments, we use ground truth boxes of people and objects as our Foreground nodes as we evaluate scene graph classification (SGCls) and

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

25

ground truth boxes of people and objects as our Foreground nodes as we evaluate scene graph classification (SGCls) and predicate classification (PredCls).Unless otherwise specified, we use SlowFast 88which corresponds to 32 input frames where the video is subsampled by a factor of 2. This means that each feature node, hi, in the graph aggregates t= 2:1seconds of temporal information on A V A as the videos are sampled at 30 frames per second (fps). As keyframes in A V A are defined at 1 fps, it means that we need to set s2keyframes for the temporal information captured by temporally-adjacent nodes in the graph to not overlap. For Action Genome, the Charades videos are sampled at 24 fps and the keyframes are on average 0.85 seconds, or 20.5 frames apart. We train our network for 20 epochs using synchronous SGD on 8 GPUs and a total batch size of 64, initialising from a Kinetics-400 [28] pretrained model. The baseline model in all our experiments is SlowFast without any graph module trained in an identical manner. When training spatio-temporal graph models, we finetune for 10 epochs

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

26

graph module trained in an identical manner. When training spatio-temporal graph models, we finetune for 10 epochs from a model trained with only a spatial graph and reduce the batch size by a factor of c. We also “freeze” batch normalisation statistics as they have high variance for small batches [19]. Full training details are in the appendix. 4.1. Experiments on video scene graph prediction Spatial message passing We first ablate only the spatial component (Sec. 3.2) of our model in Tab. 1a. Graph Attention (GAT) [57] performs slightly better than Nonlocal [56, 59] as a message passing function, and we obtain a further small improvement by combining the two methods in parallel (Sec. 3.4). Overall, we improve upon our baseline model, which is SlowFast based on ResNet 50 without any graph modelling, by 2.4 points for R@20 on SGCls. Another baseline is to insert a Non-Local layer [59] after the final res 5layer and before the ROI-Align layer, since Non-local can also be viewed as a graph network (Sec. 3.4).

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

27

Non-local can also be viewed as a graph network (Sec. 3.4). In this case, the Foreground and Context nodes are not explicitly modelled as in our method and the overall performance is 2.2 points less than our method for the R@20 of SGCls. This suggests that explicit modelling of the Foreground nodes which are subsequently classified is important to performance. We note, however, that Non-Local [59] is typically employed earlier in a network ( i.e. in res 3or res4[59, 61, 11]) and can thus be seen as a complementary Table 2: Comparison to existing methods on the Action Genome dataset [23]. Previous methods reported by [23]. SGCls PredCls R@20 R@50 R@10 R@20 MSDN [34] 44.0 47.2 – – IMP [64] 44.1 47.4 – – RelDN [69] 46.7 49.4 – – SlowFast (ResNet 50) 48.9 51.3 78.7 93.8 Ours (ResNet 50) 53.8 56.0 79.3 94.2 personsofacouch table laptopnot looking atnot contacting not contactingnot looking at not contacting in front of looking aton the side of

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

28

personsofacouch table laptopnot looking atnot contacting not contactingnot looking at not contacting in front of looking aton the side of in front of person sofacouchtelevision laptop table not looking atnot contactingin front of not contacting behindlooking at not looking atbeneathlooking at in front of sitting on on the side ofabove touching Figure 4: Example scene graph predictions for consecutive keyframes of Action Genome [23]. Classified bounding boxes are on top, and the corresponding scene graph (objects are shown as coloured rectangles, and relations with light-blue ovals) below. method to improve features learned by the network. Temporal message passing Table 1b adds temporal connections (Sec. 3.3) to our model, using GAT for message passing, as it outperformed Non-local in Tab. 1a. We observe consistent improvements for a wide range of temporal contexts,c, and strides, s, showing the utility of modelling temporal dynamics. We improve over a graph with only spatial connections by as much as 2.7 points for R@20. For a temporal model, another baseline to consider is to simply increase the temporal information available to the spatial-only graph model by increasing the number of

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

29

For a temporal model, another baseline to consider is to simply increase the temporal information available to the spatial-only graph model by increasing the number of frames at the input. When tripling the number of input frames, the R@20 is 52.4, less than all variants of our temporal model with c= 3. Similarly, increasing input frames by a factor of 5, the R@20 is 52.8, less than all our temporal models with c= 5. This improvement demonstrates the benefits of message passing on an explicit temporal graph. Iterations of message passing Table 1c shows that the model’s performance plateaus after 2 iterations of message passing, where parameters are not tied across iterations. Comparison to state-of-the-art Finally, we compare to prior works in Tab. 2 which are single-image models evalu-Table 3: Analysis of message passing neighbourhood on A V A using SlowFast with a ResNet-50 backbone as the baseline. By varying the neighbourhood, we study the effect of object representations (implicit, explicit or none) and temporal connections. We report the Frame mAP on the three types of action classes in A V A. Message passing neighbourhood PoseHumanHumanHumanObjectAll SlowFast baseline (none) 43.1 25.2 17.4 24.8

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

30

Message passing neighbourhood PoseHumanHumanHumanObjectAll SlowFast baseline (none) 43.1 25.2 17.4 24.8 Actors only 43.2 27.0 17.8 25.6 Implicit objects only 43.4 26.7 18.0 25.7 Explicit objects only 43.0 26.7 17.8 25.5 Actor + Implicit 43.4 26.8 18.3 25.9 Actor + Implicit + Explicit 43.7 27.0 18.4 26.1 Spatio-temporal 43.8 27.5 19.9 27.0 ated by [23]. Our SlowFast, ResNet 50 3D baseline outperforms these, showing the importance of using spatiotemporal features for this task. As Action Genome contains videos of humans acting to scripts, there is temporal structure in the interactions between actors and objects in the scene (Fig. 4). Our final spatio-temporal graph structured model improves substantially upon this baseline by 4.9 and 4.7 points for the R@20 and R@50 for SGCls respectively. Our improvements over the baseline for PredCls are less, as this task is easier and the performance is saturated. 4.2. Experiments on action recognition Graph structure and object representation Table 3 compares the effect of changing the neighbourhood that

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

31

this task is easier and the performance is saturated. 4.2. Experiments on action recognition Graph structure and object representation Table 3 compares the effect of changing the neighbourhood that passes messages to each foreground node that is subsequently classified. We report the performance across the three types of action categories in A V A [17] – Pose, HumanHuman and Human-Object – to show the effect that different object representations (implicit, explicit or neither as described in Sec. 3.2) and also temporal connections have. Our SlowFast baseline does not model any explicit graph and hence effectively performs no message passing. For our A V A experiments, Foreground nodes in the graph correspond to bounding boxes of the actors in the scene, using the same person detections as [61, 11]. Thus, in this section, “Foreground” and “actor” nodes are used interchangeably. We observe from the second row of Tab. 3 that passing messages only between actor nodes (and thus not modelling object interactions) provides an overall improvement of 0.8 points. The largest gain, as expected, are for HumanHuman action classes, as these are the interactions modelled by passing messages between actor nodes in the graph. When passing messages from implicit context nodes to the actors (third row), we observe an improvement in all

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

32

When passing messages from implicit context nodes to the actors (third row), we observe an improvement in all types of action classes, but primarily in Human-Human and Human-Object classes. This is because implicit context nodes encompass the entire feature map, and thus capture information about the whole scene (as also shown in Fig. 5). Modelling objects explicitly with external regional proTable 4: Comparison on A V A. We report the Mean AP using v2.1 and v2.2 annotations. All methods pretrained on Kinetics 400. “Multiscale” refers to averaging results over three scales [61]. Method v2.1 v2.2 ACRN (S3D) [54] 17.4 – Zhang et al. (R50) [71] 22.2 – SlowFast baseline (R50) 24.5 24.8 Girdhar et al. (I3D) [14] 25.0 – LFB (R50) [61] 25.8 – Ours (R50) 26.5 27.0 Ours Multiscale (R50) 27.3 27.7 SlowFast baseline (R101) 26.3 26.7 LFB (R101) [61] 26.8 – Ours (R101) 28.3 28.8

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

33

SlowFast baseline (R101) 26.3 26.7 LFB (R101) [61] 26.8 – Ours (R101) 28.3 28.8 LFB Multiscale (R101) [61] 27.7 – Ours Multiscale (R101) 29.5 30.0 posals (fourth row) improves on the same action types as the implicit object model, but performs marginally worse. This suggests that our RPN trained on OpenImages [31] is unable to detect the objects that are most discriminative of A V A actions. As object annotations are not provided in A V A, it is not possible to evaluate the recall of our external region proposals on relevant objects. It is not clear if we should expect more improvement from explicit objects, because although previous works have considered implicit [14, 54, 71] and explicit representations [3, 60], we are not aware of any that have compared the two. Note that these explicit context nodes also model Human-Human interactions as our RPN is trained on people in OpenImages [31]. We obtain further improvements by combining messages from actors, implicit and explicit context nodes, as shown by the next two rows. Finally, we evaluate our spatio-temporal model, which passes messages from actors, implicit- and explicit-context

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

34

and explicit context nodes, as shown by the next two rows. Finally, we evaluate our spatio-temporal model, which passes messages from actors, implicit- and explicit-context nodes withc= 3ands= 3, corresponding to an overall temporal window of 8.5 seconds. This model performs the best, improving upon the baseline by 2.2 points, or a relative improvement of 8.9%. In particular, the temporal connections help to improve on Human-Object action classes. A V A is a long-tailed dataset, and the most common failure modes of all variants of our model are the classes with few training examples. However, our model improves on both head and tail classes with respect to the baseline. This, and detailed per-class results, are in the appendix. State-of-the-art-comparison on A V A Table 4 compares to recent, published work on A V A. Our method builds upon SlowFast [11] as the base architecture, and our graph model shows substantial improvements with either a 3D ResNet 50 or ResNet 101 backbone. As discussed in Sec. 3.6, Girdharet al. [14] and LFB [61] can be considered as special

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

35

or ResNet 101 backbone. As discussed in Sec. 3.6, Girdharet al. [14] and LFB [61] can be considered as special cases of our graph model as they model only spatial andTable 5: Comparison to state-of-the-art on UCF101-24. We report the Frame AP at 0.5 using the corrected annotations of [51]. Method Modality Mean AP ACT [26] RGB + Flow 69.5 Song et al. [52] RGB + Flow 72.1 STEP [67] RGB + Flow 75.0 Guet al. [17] RGB + Flow 76.3 MOC [36] RGB + Flow 78.0 SlowFast (R50) RGB 76.6 SlowFast (R101) RGB 77.4 Ours (R50) RGB 78.6 Ours (R101) RGB 79.3 temporal edges respectively. Our method, which constructs a spatio-temporal graph, outperforms both. Note that we outperform LFB [61] across both ResNet 50 and ResNet 101 backbones. Zhang et al . [71] also model a spatiotemporal interactions, but employ three separate graphs with a hand-crafted aggregation function which we also outperform with our single, coherent spatio-temporal graph.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

36

with a hand-crafted aggregation function which we also outperform with our single, coherent spatio-temporal graph. State-of-the-art comparison on UCF101-24 Table 5 shows that we outperform recent, published work on UCF101-24 [53], using either a 3D ResNet 50 or ResNet 101 backbone. We do so without using optical flow as an additional input modality, showing that our network can capture temporal information without it. Moreover, the improvements obtained from our graph model are consistent with our results on the Action Genome and A V A datasets. Qualitative results We visualise the spatial and temporal messages received by an actor node in the graph, on the A V A dataset, when using GAT as the message passing function. In particular, we visualise the attention weights, ijin (9), on each implicit object node (Fig. 5) and actor in neighbouring keyframes (Fig. 6), for a given actor. We observe that the network places more weight on graph nodes that are intuitively consistent with the final action prediction. For example, in Fig. 5, the model focuses on the actor’s face when his action is “talk to” and the other person’s face when his action is “listen to”.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

37

when his action is “talk to” and the other person’s face when his action is “listen to”. 5. Conclusion and Future Work We have presented a novel spatio-temporal graph neural network framework to explicitly model interactions between actors, objects and their environment. Our formulation can model objects either implicitly or explicitly, and generalises previous structured models for video understanding [14, 54, 61, 71]. Using our versatile approach, we have achieved state-of-the-art results on two diverse tasks across three datasets. Future work remains to harness explicit object representations more effectively on A V A. (a)talk to (0.6), sit (0.92), watch person (0.81) (b)listen to person (0.9), watch person (0.89), sit (0.97) (c)carry/hold (1.0), sit(0.88) Figure 5: Visualisation of spatial message passing when using implicit objects as detailed in Sec. 4.2. Note in (a) and (b) that the network focuses on the selected actor (denoted by the green box on the left) when the action is “talk to”, and on the person speaking to him when

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

38

focuses on the selected actor (denoted by the green box on the left) when the action is “talk to”, and on the person speaking to him when the action is “listen to person”. Attention weights, ijare colour-coded according to the bar on the far right. Best viewed in colour. talk to (0.85), watch a person (0.82), sit(0.51) Figure 6: Visualisation of temporal graph as detailed in Sec. 4.2. We show the neighbours for the actor depicted by the red box in the centre keyframe, colour-coding the respective boxes with their attention weights ( ij). Predicted action scores of the red actor are below, and selected keyframes are s= 2:1seconds apart. Notice that the model focuses on the people 2 keyframes (4.2 seconds) on either side of the actor to recognise that she is “talking to” and “watching” them. References [1] Anurag Arnab, Chen Sun, Arsha Nagrani, and Cordelia Schmid. Uncertainty-aware weakly supervised action detection from untrimmed videos. In ECCV , 2020. 2 [2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

39

[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. In arXiv preprint arXiv:1607.06450 , 2016. 5 [3] Fabien Baradel, Natalia Neverova, Christian Wolf, Julien Mille, and Greg Mori. Object level visual reasoning in videos. In ECCV , 2018. 1, 2, 4, 8 [4] Peter Battaglia, Razvan Pascanu, Matthew Lai, and Danilo Jimenez Rezende. Interaction networks for learning about objects, relations and physics. In NeurIPS , 2016. 3, 4, 5 [5] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. In arXiv preprint arXiv:1806.01261 , 2018. 2, 3, 5 [6] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR , 2017. 2

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

40

recognition? a new model and the kinetics dataset. In CVPR , 2017. 2 [7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV , 2018. 1 [8] Yunpeng Chen, Yannis Kalantidis, Jianshu Li, Shuicheng Yan, and Jiashi Feng. Aˆ 2-nets: Double attention networks. InNeurIPS , 2018. 2 [9] Yunpeng Chen, Marcus Rohrbach, Zhicheng Yan, Yan Shuicheng, Jiashi Feng, and Yannis Kalantidis. Graph-based global reasoning networks. In CVPR , 2019. 2 [10] Bo Dai, Yuqi Zhang, and Dahua Lin. Detecting visual relationships with deep relational networks. In CVPR , 2017. 3[11] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV , 2019. 6, 7, 8, 12 [12] Christoph Feichtenhofer, Axel Pinz, and Richard Wildes.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

41

ICCV , 2019. 6, 7, 8, 12 [12] Christoph Feichtenhofer, Axel Pinz, and Richard Wildes. Spatiotemporal residual networks for video action recognition. In NeurIPS , 2016. 2 [13] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICML , 2017. 1, 2, 3, 5 [14] Rohit Girdhar, Joao Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR , 2019. 1, 2, 3, 5, 8 [15] Ross Girshick. Fast r-cnn. In ICCV , 2015. 6 [16] Georgia Gkioxari, Ross Girshick, and Jitendra Malik. Contextual action recognition with r* cnn. In ICCV , 2015. 2 [17] Chunhui Gu, Chen Sun, David A Ross, Carl V ondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

42

George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In CVPR , 2018. 1, 2, 4, 6, 7, 8, 12 [18] Abhinav Gupta, Aniruddha Kembhavi, and Larry S Davis. Observing human-object interactions: Using spatial and functional compatibility for recognition. PAMI , 31(10), 2009. 2 [19] Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Girshick. Mask r-cnn. In ICCV , 2017. 4, 6 [20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR , 2016. 1, 6 [21] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR , 2018. 2 [22] Ashesh Jain, Amir R Zamir, Silvio Savarese, and Ashutosh Saxena. Structural-rnn: Deep learning on spatio-temporal graphs. In CVPR , 2016. 2 [23] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

43

graphs. In CVPR , 2016. 2 [23] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In CVPR , 2020. 2, 3, 4, 6, 7, 12 [24] Huaizu Jiang, Ishan Misra, Marcus Rohrbach, Erik LearnedMiller, and Xinlei Chen. In defense of grid features for visual question answering. In CVPR , 2020. 4 [25] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In CVPR , 2015. 3 [26] Vicky Kalogeiton, Philippe Weinzaepfel, Vittorio Ferrari, and Cordelia Schmid. Action tubelet detector for spatiotemporal action localization. In ICCV , 2017. 2, 8 [27] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR , 2014. 2 [28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang,

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

44

2014. 2 [28] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. In arXiv preprint arXiv:1705.06950 , 2017. 6 [29] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR , 2017. 2, 3, 5 [30] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV , 123(1), 2017. 3 [31] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

45

Popov, Matteo Malloci, Tom Duerig, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV , 2020. 4, 8 [32] Ivan Laptev. On space-time interest points. IJCV , 64(2-3), 2005. 2 [33] Yikang Li, Wanli Ouyang, Xiaogang Wang, and Xiao’ou Tang. Vip-cnn: Visual phrase guided convolutional neural network. In CVPR , 2017. 3 [34] Yikang Li, Wanli Ouyang, Bolei Zhou, Kun Wang, and Xiaogang Wang. Scene graph generation from objects, phrases and region captions. In ICCV , 2017. 7 [35] Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. In ICLR , 2016. 5 [36] Yixuan Li, Zixu Wang, Limin Wang, and Gangshan Wu. Actions as moving points. In ECCV , 2020. 8 [37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays,

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

46

[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV , 2014. 12 [38] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV , 2016. 2 [39] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li FeiFei. Visual relationship detection with language priors. In ECCV , 2016. 6 [40] Chih-Yao Ma, Asim Kadav, Iain Melvin, Zsolt Kira, Ghassan AlRegib, and Hans Peter Graf. Attend and interact: Higher-order object interactions for video understanding. In CVPR , 2018. 2 [41] Marcin Marszalek, Ivan Laptev, and Cordelia Schmid. Actions in Context. In CVPR , 2009. 2 [42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu,

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

47

[42] Joanna Materzynska, Tete Xiao, Roei Herzig, Huijuan Xu, Xiaolong Wang, and Trevor Darrell. Something-else: Compositional action recognition with spatial-temporal interaction networks. In CVPR , 2020. 2 [43] Aude Oliva and Antonio Torralba. The role of context in object recognition. Trends in cognitive sciences , 11(12), 2007. 2 [44] Siyuan Qi, Wenguan Wang, Baoxiong Jia, Jianbing Shen, and Song-Chun Zhu. Learning human-object interactions by graph parsing neural networks. In ECCV , 2018. 2 [45] Andrew Rabinovich, Andrea Vedaldi, Carolina Galleguillos, Eric Wiewiora, and Serge Belongie. Objects in context. In ICCV , 2007. 2 [46] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS , 2015. 1, 2, 3, 4, 12 [47] Adam Santoro, David Raposo, David G Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

48

[47] Adam Santoro, David Raposo, David G Barrett, Mateusz Malinowski, Razvan Pascanu, Peter Battaglia, and Timothy Lillicrap. A simple neural network module for relational reasoning. In NeurIPS , 2017. 5 [48] Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks , 20(1), 2008. 2, 5 [49] Gunnar A. Sigurdsson, G. Varol, X. Wang, Ali Farhadi, I. Laptev, and A. Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV , 2016. 1, 3, 6 [50] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS , 2014. 2 [51] Gurkirt Singh, Suman Saha, Michael Sapienza, Philip HS Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV , 2017. 2, 8

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

49

Torr, and Fabio Cuzzolin. Online real-time multiple spatiotemporal action localisation and prediction. In ICCV , 2017. 2, 8 [52] Lin Song, Shiwei Zhang, Gang Yu, and Hongbin Sun. Tacnet: Transition-aware context network for spatio-temporal action detection. In CVPR , 2019. 8 [53] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. In arXiv preprint arXiv:1212.0402 , 2012. 2, 6, 8, 12 [54] Chen Sun, Abhinav Shrivastava, Carl V ondrick, Kevin Murphy, Rahul Sukthankar, and Cordelia Schmid. Actor-centric relation network. In ECCV , 2018. 1, 2, 3, 4, 5, 8 [55] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR , 2018. 2

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

50

LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR , 2018. 2 [56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS , 2017. 2, 5, 6, 13 [57] Petar Veli ˇckovi ´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. In ICLR , 2018. 3, 5, 6, 12 [58] Heng Wang, Alexander Kl ¨aser, Cordelia Schmid, and Cheng-Lin Liu. Dense trajectories and motion boundary descriptors for action recognition. IJCV , 103(1), 2013. 2 [59] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR , 2018. 2, 5, 6, 12 [60] Xiaolong Wang and Abhinav Gupta. Videos as space-time

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

51

12 [60] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In ECCV , 2018. 1, 2, 3, 4, 5, 8 [61] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In CVPR , 2019. 1, 2, 3, 5, 6, 7, 8, 12 [62] Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. Pay less attention with lightweight and dynamic convolutions. In ICLR , 2019. 2 [63] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github. com/facebookresearch/detectron2 , 2019. 12 [64] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In CVPR , 2017. 3, 7 [65] Huijuan Xu, Lizhi Yang, Stan Sclaroff, Kate Saenko, and

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

52

CVPR , 2017. 3, 7 [65] Huijuan Xu, Lizhi Yang, Stan Sclaroff, Kate Saenko, and Trevor Darrell. Spatio-temporal action detection with multiobject interaction. In arXiv preprint arXiv:2004.00180 , 2020. 2 [66] Jianwei Yang, Jiasen Lu, Stefan Lee, Dhruv Batra, and Devi Parikh. Graph r-cnn for scene graph generation. In ECCV , 2018. 3 [67] Xitong Yang, Xiaodong Yang, Ming-Yu Liu, Fanyi Xiao, Larry S Davis, and Jan Kautz. Step: Spatio-temporal progressive learning for video action detection. In CVPR , 2019. 8 [68] Bangpeng Yao and Li Fei-Fei. Modeling mutual context of object and human pose in human-object interaction activities. In CVPR , 2010. 2 [69] Ji Zhang, Kevin J Shih, Ahmed Elgammal, Andrew Tao, and Bryan Catanzaro. Graphical contrastive losses for scene graph parsing. In CVPR , 2019. 7 [70] Li Zhang, Dan Xu, Anurag Arnab, and Philip HS Torr. Dynamic graph message passing networks. In CVPR , 2020. 2

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

53

[70] Li Zhang, Dan Xu, Anurag Arnab, and Philip HS Torr. Dynamic graph message passing networks. In CVPR , 2020. 2 [71] Yubo Zhang, Pavel Tokmakov, Martial Hebert, and Cordelia Schmid. A structured model for action detection. In CVPR , 2019. 1, 2, 3, 5, 8 [72] Bolei Zhou, Alex Andonian, Aude Oliva, and Antonio Torralba. Temporal relational reasoning in videos. In ECCV , 2018. 2 Appendix In this appendix, we additional analysis on A V A (Sec. A) and implementation details (Sec. B). A. A V A per-class analysis Figure 7 presents per-class results on A V A for our SlowFast baseline, spatial and spatio-temporal graph models when using a ResNet-50 backbone. We additionally show the number of examples in the training set. Our proposed graph models improve on both head and tail classes in A V A. In fact, the largest absolute improvements are observed on tail classes such as “cut”, “swim” and “sing to”. However, note that the absolute accuracy of each class is still correlated with the number of training examples. The classes which all variants of our model perform the worst on, are

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

54

note that the absolute accuracy of each class is still correlated with the number of training examples. The classes which all variants of our model perform the worst on, are the classes with the least training examples such as “point to (an object)”, “take a photo” and “turn (eg a screwdriver)”. To emphasise the long-tailed distribution of A V A, note that “point to (an object)” has the fewest training examples (just 97), whilst “watch (a person)” has the most with 168 148. B. Additional Implementation details Person detections For our spatio-temporal action detection experiments, we use external person detections as our actor region proposals like [11, 61]. For our experiments on A V A, we use the person detections publicly released by [11, 61] during both training and testing. More specifically, this is a Faster-RCNN [46] model, pretrained on Microsoft COCO [37] and then finetuned on the training set of A V A [17], using Detectron [63]. For our experiments on UCF101-24 [53], we finetune a Faster-RCNN detector pretrained on COCO on the UCF101-24 training set using Detectron.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

55

UCF101-24 training set using Detectron. When training on both A V A and UCF101-24 datasets, we use both ground-truth and predicted person bounding boxes. Predicted bounding boxes are assigned labels by matching them to ground-truth boxes using an IoU threshold of 0.75. Predicted bounding boxes which do not match any ground truth boxes act as negative examples for all action classes. OpenImages object detector For our “explicit object representation” experiments on A V A, we use the publicly available Faster-RCNN detector trained on OpenImages v41. Optimiser hyperparameters We train our models following the settings of the publicly released SlowFast code [11]. We train for 20 epochs using synchronous Stochastic Gradient Descent (SGD) and a momentum of 0.9 1https://tfhub.dev/google/faster_rcnn/ openimages_v4/inception_resnet_v2/1on 8 Nvidia V100 (16 GB) GPUs. The learning rate was set to 0.1, and reduced by a factor of 10 after 10 and 15 epochs respectively. We employed a linear-warmup schedule for the learning rate, increasing linearly from 1:25104to0:1

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

56

respectively. We employed a linear-warmup schedule for the learning rate, increasing linearly from 1:25104to0:1 in the first 5 epochs. We also used a weight decay of 107. An epoch is defined as all the keyframes in the dataset. Training loss functions For spatio-temporal action recognition on A V A and UCF101-24, we use the binary cross-entropy loss function. The loss function for a single example is: L(x;y) =1 CCX iyilog ((xi))+(1yi) log (1(xi)); (10) whereCis the number of classes, is the sigmoid activation function, x2RCdenotes the logits predicted by the network and yi2f0;1gdenotes the binary ground truth for theithclass. For scene graph classification, we use two loss functions: One for predicting the object class of each bounding box proposal, and another for predicting the relationship label between each of theN(N1) 2pairs of object proposals. For the former, we use the softmax cross-entropy, as each proposal can only be assigned a single object class. And for

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

57

2pairs of object proposals. For the former, we use the softmax cross-entropy, as each proposal can only be assigned a single object class. And for the latter, we use the sigmoid cross-entropy, as there can be multiple relationship labels between any pair of objects. Concretely, the loss is L(x;w;y;z ) =Lobject(x;y) +Lrel(w;z) (11) Lobject(x;y) =1 NNX iCX jyijlog (Softmax (x)ij)(12) Lrel(w;z) =2 N(N1)RNX ii1X jRX rzijrlog ((wijr)) + (1zijr) log (1(wijr)): (13) Here,Nis the number of object proposals in the video clip, Rthe number of relationship classes and Cthe number of object classes. yis the one-hot, ground-truth object class label,zthe binary ground-truth relationship label, and xand wdenote the object- and relationship-logits respectively. On Action Genome [23], there are R= 25 relationship

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

58

wdenote the object- and relationship-logits respectively. On Action Genome [23], there are R= 25 relationship classes, and C= 35 object classes. We set = 0:5during training on Action Genome to prevent the object class loss from dominating the overall loss (the softmax cross entropy has a higher loss than the binary sigmoid cross entropy at the start of training). Multi-headed attention For the GAT [57] and NonLocal [59] message-passing functions used in the paper Mean AP Number of training examples 0.000.250.500.751.00 100100010000100000watch (a person) stand talk to (e.g. self a person a listen to (a person) sit carry/hold (an object) walk touch (an object) bend/bow (at the waist) lie/sleep ride (e.g. a bike a car a horse) run/jog answer phone dance eat smoke fight/hit (a person) drink crouch/kneel read martial art grab (a person) watch (e.g. TV) sing to (e.g. self a person a play musical instrument

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

59

read martial art grab (a person) watch (e.g. TV) sing to (e.g. self a person a play musical instrument open (e.g. a window a car drive (e.g. a car a truck) hand clap get up hug (a person) give/serve (an object) to (a close (e.g. a door a box) write sail boat kiss (a person) listen (e.g. to music) hand shake take (an object) from (a put down lift/pick up text on/look at a cellphone lift (a person) push (an object) push (another person) hand wave pull (an object) dress/put on clothing fall down climb (e.g. a mountain) throw jump/leap enter shoot cut take a photo hit (an object) work on a computer turn (e.g. a screwdriver) swim point to (an object) Baseline Spatial Spatio-temporal Training examplesFigure 7: Mean Average Precision for each action class on A V A, for the SlowFast baseline, spatial- and spatio-temporal graph models with

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15662

60

a ResNet-50 backbone. The number of examples in the training set are shown with the blue line on the right, vertical axis. Note that a logarithmic scale is used, as A V A has a long-tailed distribution of labels. Table 6: Ablation study of the number of heads when using GAT as the message passing function. Mean AP reported on A V A using a ResNet-50 backbone. Number of heads 1 2 3 4 5 Mean AP 25.4 25.8 25.9 25.9 25.9 (Sec 3.4), it is common to compute “multi-headed” attention [56]. In all experiments reported in the main paper, we used 4 heads, as motivated by the ablation study using GAT on the A V A dataset in Tab. 6. To combine the messages from each head, we performed an attention-weighted convex combination as done in Eq. 9. Multiscale testing We use three scales when performing multiscale testing (as done in Table 3 of the main paper). Specifically, we resize input video-clips so that the shortest side is 224, 256 and 320 respectively.

2103.15662

Unified Graph Structured Models for Video Understanding

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model's performance. We demonstrate our method on two different tasks requiring relational reasoning in videos -- spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset -- and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

http://arxiv.org/pdf/2103.15662

cs.CV

20210329

20210329

2103.15691

0

ViViT: A Video Vision Transformer Anurag Arnab*Mostafa Dehghani*Georg Heigold Chen Sun Mario Lu ˇci´c†Cordelia Schmid† Google Research faarnab, dehghani, heigold, chensun, lucic, cordelias [email protected] Abstract We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatiotemporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

1

video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic. 1. Introduction Approaches based on deep convolutional neural networks have advanced the state-of-the-art across many standard datasets for vision problems since AlexNet [38]. At the same time, the most prominent architecture of choice in sequence-to-sequence modelling (e.g. in natural language processing) is the transformer [68], which does not use convolutions, but is based on multi-headed self-attention. This operation is particularly effective at modelling long-range dependencies and allows the model to attend over all elements in the input sequence. This is in stark contrast to convolutions where the corresponding “receptive field” is limited, and grows linearly with the depth of the network. The success of attention-based models in NLP has recently inspired approaches in computer vision to integrate transformers into CNNs [75, 7], as well as some attempts to replace convolutions completely [49, 3, 53]. However, it is *Equal contribution

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

2

transformers into CNNs [75, 7], as well as some attempts to replace convolutions completely [49, 3, 53]. However, it is *Equal contribution †Equal advisingonly very recently with the Vision Transformer (ViT) [18], that a pure-transformer based architecture has outperformed its convolutional counterparts in image classification. Dosovitskiy et al. [18] closely followed the original transformer architecture of [68], and noticed that its main benefits were observed at large scale – as transformers lack some of the inductive biases of convolutions (such as translational equivariance), they seem to require more data [18] or stronger regularisation [64]. Inspired by ViT, and the fact that attention-based architectures are an intuitive choice for modelling longrange contextual relationships in video, we develop several transformer-based models for video classification. Currently, the most performant models are based on deep 3D convolutional architectures [8, 20, 21] which were a natural extension of image classification CNNs [27, 60]. Recently, these models were augmented by incorporating selfattention into their later layers to better capture long-range dependencies [75, 23, 79, 1].

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

3

dependencies [75, 23, 79, 1]. As shown in Fig. 1, we propose pure-transformer models for video classification. The main operation performed in this architecture is self-attention, and it is computed on a sequence of spatio-temporal tokens that we extract from the input video. To effectively process the large number of spatio-temporal tokens that may be encountered in video, we present several methods of factorising our model along spatial and temporal dimensions to increase efficiency and scalability. Furthermore, to train our model effectively on smaller datasets, we show how to reguliarise our model during training and leverage pretrained image models. We also note that convolutional models have been developed by the community for several years, and there are thus many “best practices” associated with such models. As pure-transformer models present different characteristics, we need to determine the best design choices for such architectures. We conduct a thorough ablation analysis of tokenisation strategies, model architecture and regularisation methods. Informed by this analysis, we achieve stateof-the-art results on multiple standard video classification benchmarks, including Kinetics 400 and 600 [35], Epic

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

4

benchmarks, including Kinetics 400 and 600 [35], Epic Kitchens 100 [13], Something-Something v2 [26] and Moments in Time [45].arXiv:2103.15691v2 [cs.CV] 1 Nov 2021 …1CLS032NPosition + Token EmbeddingMLP HeadClassFactorisedEncoderL× K V Q Self-AttentionTransformer Encoder MLPLayer Norm Layer NormMulti-HeadDot-Product Attention Embed to tokensFactorisedSelf-Attention 21NFactorisedDot-Product ●●●SpatialSpatial●●●TemporalTemporal SpatialTemporalSpatialTemporal●●●SpatialTemporal●●●FuseSpatialTemporalFuse ●●●21N●●●21N●●●Figure 1: We propose a pure-transformer architecture for video classification, inspired by the recent success of such models for images [18]. To effectively process a large number of spatio-temporal tokens, we develop several model variants which factorise different components of the transformer encoder over the spatial- and temporal-dimensions. As shown on the right, these factorisations correspond to different attention patterns over space and time. 2. Related Work

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

5

attention patterns over space and time. 2. Related Work Architectures for video understanding have mirrored advances in image recognition. Early video research used hand-crafted features to encode appearance and motion information [41, 69]. The success of AlexNet on ImageNet [38, 16] initially led to the repurposing of 2D image convolutional networks (CNNs) for video as “twostream” networks [34, 56, 47]. These models processed RGB frames and optical flow images independently before fusing them at the end. Availability of larger video classification datasets such as Kinetics [35] subsequently facilitated the training of spatio-temporal 3D CNNs [8, 22, 65] which have significantly more parameters and thus require larger training datasets. As 3D convolutional networks require significantly more computation than their image counterparts, many architectures factorise convolutions across spatial and temporal dimensions and/or use grouped convolutions [59, 66, 67, 81, 20]. We also leverage factorisation of the spatial and temporal dimensions of videos to increase efficiency, but in the context of transformer-based models. Concurrently, in natural language processing (NLP),

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

6

of the spatial and temporal dimensions of videos to increase efficiency, but in the context of transformer-based models. Concurrently, in natural language processing (NLP), Vaswani et al. [68] achieved state-of-the-art results by replacing convolutions and recurrent networks with the transformer network that consisted only of self-attention, layer normalisation and multilayer perceptron (MLP) operations. Current state-of-the-art architectures in NLP [17, 52] remain transformer-based, and have been scaled to web-scale datasets [5]. Many variants of the transformer have also been proposed to reduce the computational cost of selfattention when processing longer sequences [10, 11, 37, 62, 63, 73] and to improve parameter efficiency [40, 14]. Although self-attention has been employed extensively in computer vision, it has, in contrast, been typically incorporated as a layer at the end or in the later stages of the network [75, 7, 32, 77, 83] or to augment residualblocks [30, 6, 9, 57] within a ResNet architecture [27].

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

7

the network [75, 7, 32, 77, 83] or to augment residualblocks [30, 6, 9, 57] within a ResNet architecture [27]. Although previous works attempted to replace convolutions in vision architectures [49, 53, 55], it is only very recently that Dosovitisky et al. [18] showed with their ViT architecture that pure-transformer networks, similar to those employed in NLP, can achieve state-of-the-art results for image classification too. The authors showed that such models are only effective at large scale, as transformers lack some of inductive biases of convolutional networks (such as translational equivariance), and thus require datasets larger than the common ImageNet ILSRVC dataset [16] to train. ViT has inspired a large amount of follow-up work in the community, and we note that there are a number of concurrent approaches on extending it to other tasks in computer vision [71, 74, 84, 85] and improving its dataefficiency [64, 48]. In particular, [4, 46] have also proposed transformer-based models for video. In this paper, we develop pure-transformer architectures for video classification. We propose several variants of our

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101

2103.15691

8

transformer-based models for video. In this paper, we develop pure-transformer architectures for video classification. We propose several variants of our model, including those that are more efficient by factorising the spatial and temporal dimensions of the input video. We also show how additional regularisation and pretrained models can be used to combat the fact that video datasets are not as large as their image counterparts that ViT was originally trained on. Furthermore, we outperform the stateof-the-art across five popular datasets. 3. Video Vision Transformers We start by summarising the recently proposed Vision Transformer [18] in Sec. 3.1, and then discuss two approaches for extracting tokens from video in Sec. 3.2. Finally, we develop several transformer-based architectures for video classification in Sec. 3.3 and 3.4. 3.1. Overview of Vision Transformers (ViT) Vision Transformer (ViT) [18] adapts the transformer architecture of [68] to process 2D images with minimal changes. In particular, ViT extracts Nnon-overlapping image patches, xi2Rhw, performs a linear projection and then rasterises them into 1D tokens zi2Rd. The sequence

2103.15691

ViViT: A Video Vision Transformer

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatio-temporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks. To facilitate further research, we release code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

http://arxiv.org/pdf/2103.15691

ICCV 2021. Code at https://github.com/google-research/scenic/tree/main/scenic/projects/vivit

cs.CV

20210329

20211101