|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- laion/laion400m |
|
- kakaobrain/coyo-700m |
|
pipeline_tag: feature-extraction |
|
tags: |
|
- Vision |
|
- LLaVA |
|
--- |
|
|
|
|
|
|
|
|
|
[[Paper]](https://arxiv.org/abs/2407.17331) [[GitHub]](https://github.com/deepglint/unicom) |
|
## Model |
|
We used the same Vision Transformer architecture [ViT-L/14@336px as CLIP](https://huggingface.co/openai/clip-vit-large-patch14-336). |
|
|
|
 |
|
|
|
|
|
## Data |
|
Our model was trained on publicly available image-caption data from the [LAION400M](https://arxiv.org/abs/2111.02114) and [COYO700M](https://github.com/kakaobrain/coyo-dataset) datasets. |
|
|
|
## Performance and Limitations |
|
|
|
### A. MLLMs Evaluation Results |
|
In our experiments, we replaced the CLIP model in [LLaVA-NeXT](https://github.com/LLaVA-VL/LLaVA-NeXT) with the MLCD model to demonstrate the performance of the MLCD model in Multimodal Large Language Models (MLLMs). For the language model, we used [Qwen2.5-7B](https://huggingface.co/Qwen/Qwen2.5-7B). The evaluation results show that the modified model performs exceptionally well across multiple benchmarks, validating the effectiveness of the MLCD model within MLLMs. |
|
|
|
| Vision Tower | MLCD (ViT_L_14_336px) | CLIP (ViT_L_14_336px) | |
|
|:----------------|:-------------|:-------------| |
|
| LLM | Qwen2.5-7B | Qwen2.5-7B | |
|
| AI2D | **76.98** | 73.15 | |
|
| ScienceQA_img | **78.09** | 76.35 | |
|
| GQA | **64.17** | 63.31 | |
|
| InfoVQA_val | **43.48** | 38.88 | |
|
| MMBench_cn_dev | **74.83** | 72.51 | |
|
| MMBench_en_dev | **76.37** | 74.57 | |
|
| MME(cognition) | **432** | 384 | |
|
| MME(perception) | **1598** | 1512 | |
|
| SeedBench | **68.20** | 66.80 | |
|
| SeedBench_img | **73.75** | 72.72 | |
|
| MMStar | **50.98** | 48.98 | |
|
| MMMU | **44.30** | 44.20 | |
|
| OCRBench | **531.00** | 525.00 | |
|
| ChartQA | **67.84** | 66.52 | |
|
| DocVQA_val | **76.46** | 75.21 | |
|
| POPE | 88.69 | **88.83** | |
|
| TextVQA_val | 61.69 | **62.47** | |
|
|
|
|
|
|
|
|
|
### B. Linear Probe Evaluation Results |
|
This table presents the results of linear probe evaluations comparing CLIP and MLCD models on the ViT_L_14_336px architecture across various datasets. The linear probe test freezes the pre-trained model's weights and trains a linear classifier on top to assess how well the model's representations generalize to different tasks. |
|
|
|
| Dataset | MLCD (ViT_L_14_336px) | CLIP (ViT_L_14_336px) | |
|
|:---------------|:----------------------|:----------------------| |
|
| **AVG** | **87.15** | 85.35 | |
|
| Food101 | **96.21** | 95.90 | |
|
| CIFAR-10 | **99.36** | 97.90 | |
|
| CIFAR-100 | **93.69** | 87.40 | |
|
| Birdsnap | **88.18** | 79.90 | |
|
| SUN397 | **87.96** | 82.20 | |
|
| Stanford Cars | **95.16** | 91.50 | |
|
| FGVC Aircraft | **86.38** | 71.60 | |
|
| Describable Textures Dataset | **86.70** | 83.00 | |
|
| Oxford-IIIT Pets | **96.27** | 95.10 | |
|
| Caltech-101 | **97.92** | 96.00 | |
|
| Flowers102 | **99.58** | 99.20 | |
|
| MNIST | 98.67 | **99.20** | |
|
| STL-10 | 99.28 | **99.70** | |
|
| EuroSAT | **99.06** | 98.10 | |
|
| RESISC45 | **95.48** | 94.90 | |
|
| GTSRB | 92.32 | **92.40** | |
|
| KITTI | **75.39** | 69.20 | |
|
| Country211 | 38.12 | **46.40** | |
|
| PatchCamelyon | **88.00** | 85.60 | |
|
| UCF101 | **92.86** | 92.00 | |
|
| Kinetics-700 | **73.35** | 73.00 | |
|
| CLEVR | **64.40** | 60.30 | |
|
| Hateful Memes | 72.00 | **77.30** | |
|
| SST-2 | 76.33 | **80.50** | |
|
| ImageNet | **86.10** | 85.40 | |
|
|
|
|
|
### C. Limitations |
|
|
|
Models with higher resolution are more friendly to OCR results. We are currently training such models and will soon make them available. |
|
|
|
|
|
## Acknowledgments |
|
|
|
We would like to express our gratitude to [Xie Yin](https://huggingface.co/Yin-Xie) and [Yumeng Wang](https://huggingface.co/devymex) for their significant contributions to the experimental validation in MLLMs. |
