README.md · qninhdt/MiDaS at main

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer

This repository contains code to compute depth from a single image. It accompanies our paper:

Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-shot Cross-dataset Transfer
René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, Vladlen Koltun

and our preprint:

Vision Transformers for Dense Prediction
René Ranftl, Alexey Bochkovskiy, Vladlen Koltun

For the latest release MiDaS 3.1, a technical report and video are available.

MiDaS was trained on up to 12 datasets (ReDWeb, DIML, Movies, MegaDepth, WSVD, TartanAir, HRWSI, ApolloScape, BlendedMVS, IRS, KITTI, NYU Depth V2) with multi-objective optimization. The original model that was trained on 5 datasets (MIX 5 in the paper) can be found here. The figure below shows an overview of the different MiDaS models; the bubble size scales with number of parameters.

Setup

Pick one or more models and download the corresponding weights to the weights folder:

MiDaS 3.1

For highest quality: dpt_beit_large_512
For moderately less quality, but better speed-performance trade-off: dpt_swin2_large_384
For embedded devices: dpt_swin2_tiny_256, dpt_levit_224
For inference on Intel CPUs, OpenVINO may be used for the small legacy model: openvino_midas_v21_small .xml, .bin

MiDaS 3.0: Legacy transformer models dpt_large_384 and dpt_hybrid_384

MiDaS 2.1: Legacy convolutional models midas_v21_384 and midas_v21_small_256

Set up dependencies:

conda env create -f environment.yaml
conda activate midas-py310

optional

For the Next-ViT model, execute

git submodule add https://github.com/isl-org/Next-ViT midas/external/next_vit

For the OpenVINO model, install

pip install openvino

Usage

Place one or more input images in the folder input.
Run the model with
```
python run.py --model_type <model_type> --input_path input --output_path output
```
where <model_type> is chosen from dpt_beit_large_512, dpt_beit_large_384, dpt_beit_base_384, dpt_swin2_large_384, dpt_swin2_base_384, dpt_swin2_tiny_256, dpt_swin_large_384, dpt_next_vit_large_384, dpt_levit_224, dpt_large_384, dpt_hybrid_384, midas_v21_384, midas_v21_small_256, openvino_midas_v21_small_256.
The resulting depth maps are written to the output folder.

optional

By default, the inference resizes the height of input images to the size of a model to fit into the encoder. This size is given by the numbers in the model names of the accuracy table. Some models do not only support a single inference height but a range of different heights. Feel free to explore different heights by appending the extra command line argument --height. Unsupported height values will throw an error. Note that using this argument may decrease the model accuracy.
By default, the inference keeps the aspect ratio of input images when feeding them into the encoder if this is supported by a model (all models except for Swin, Swin2, LeViT). In order to resize to a square resolution, disregarding the aspect ratio while preserving the height, use the command line argument --square.

via Camera

If you want the input images to be grabbed from the camera and shown in a window, leave the input and output paths away and choose a model type as shown above:

python run.py --model_type <model_type> --side

The argument --side is optional and causes both the input RGB image and the output depth map to be shown side-by-side for comparison.

via Docker

Make sure you have installed Docker and the NVIDIA Docker runtime.
Build the Docker image:
```
docker build -t midas .
```
Run inference:
```
docker run --rm --gpus all -v $PWD/input:/opt/MiDaS/input -v $PWD/output:/opt/MiDaS/output -v $PWD/weights:/opt/MiDaS/weights midas
```
This command passes through all of your NVIDIA GPUs to the container, mounts the input and output directories and then runs the inference.

via PyTorch Hub

The pretrained model is also available on PyTorch Hub

via TensorFlow or ONNX

See README in the tf subdirectory.

Currently only supports MiDaS v2.1.

via Mobile (iOS / Android)

See README in the mobile subdirectory.

via ROS1 (Robot Operating System)

See README in the ros subdirectory.

Currently only supports MiDaS v2.1. DPT-based models to be added.

Accuracy

We provide a zero-shot error $\epsilon_d$ which is evaluated for 6 different datasets (see paper). Lower error values are better. $\color{green}{\textsf{Overall model quality is represented by the improvement}}$ (Imp.) with respect to MiDaS 3.0 DPT_L-384. The models are grouped by the height used for inference, whereas the square training resolution is given by the numbers in the model names. The table also shows the number of parameters (in millions) and the frames per second for inference at the training resolution (for GPU RTX 3090):

MiDaS Model	DIW ^WHDR	Eth3d ^AbsRel	Sintel ^AbsRel	TUM ^δ1	KITTI ^δ1	NYUv2 ^δ1	$\color{green}{\textsf{Imp.}}$ ^%	Par. ^M	FPS
Inference height 512
v3.1 BEiT_L-512	0.1137	0.0659	0.2366	6.13	11.56*	1.86*	$\color{green}{\textsf{19}}$	345	5.7
v3.1 BEiT_L-512$\tiny{\square}$	0.1121	0.0614	0.2090	6.46	5.00*	1.90*	$\color{green}{\textsf{34}}$	345	5.7

Inference height 384
v3.1 BEiT_L-512	0.1245	0.0681	0.2176	6.13	6.28*	2.16*	$\color{green}{\textsf{28}}$	345	12
v3.1 Swin2_L-384$\tiny{\square}$	0.1106	0.0732	0.2442	8.87	5.84*	2.92*	$\color{green}{\textsf{22}}$	213	41
v3.1 Swin2_B-384$\tiny{\square}$	0.1095	0.0790	0.2404	8.93	5.97*	3.28*	$\color{green}{\textsf{22}}$	102	39
v3.1 Swin_L-384$\tiny{\square}$	0.1126	0.0853	0.2428	8.74	6.60*	3.34*	$\color{green}{\textsf{17}}$	213	49
v3.1 BEiT_L-384	0.1239	0.0667	0.2545	7.17	9.84*	2.21*	$\color{green}{\textsf{17}}$	344	13
v3.1 Next-ViT_L-384	0.1031	0.0954	0.2295	9.21	6.89*	3.47*	$\color{green}{\textsf{16}}$	72	30
v3.1 BEiT_B-384	0.1159	0.0967	0.2901	9.88	26.60*	3.91*	$\color{green}{\textsf{-31}}$	112	31
v3.0 DPT_L-384	0.1082	0.0888	0.2697	9.97	8.46	8.32	$\color{green}{\textsf{0}}$	344	61
v3.0 DPT_H-384	0.1106	0.0934	0.2741	10.89	11.56	8.69	$\color{green}{\textsf{-10}}$	123	50
v2.1 Large₃₈₄	0.1295	0.1155	0.3285	12.51	16.08	8.71	$\color{green}{\textsf{-32}}$	105	47

Inference height 256
v3.1 Swin2_T-256$\tiny{\square}$	0.1211	0.1106	0.2868	13.43	10.13*	5.55*	$\color{green}{\textsf{-11}}$	42	64
v2.1 Small₂₅₆	0.1344	0.1344	0.3370	14.53	29.27	13.43	$\color{green}{\textsf{-76}}$	21	90

Inference height 224
v3.1 LeViT₂₂₄$\tiny{\square}$	0.1314	0.1206	0.3148	18.21	15.27*	8.64*	$\color{green}{\textsf{-40}}$	51	73

* No zero-shot error, because models are also trained on KITTI and NYU Depth V2
$\square$ Validation performed at square resolution, either because the transformer encoder backbone of a model does not support non-square resolutions (Swin, Swin2, LeViT) or for comparison with these models. All other validations keep the aspect ratio. A difference in resolution limits the comparability of the zero-shot error and the improvement, because these quantities are averages over the pixels of an image and do not take into account the advantage of more details due to a higher resolution.
Best values per column and same validation height in bold

Improvement

The improvement in the above table is defined as the relative zero-shot error with respect to MiDaS v3.0 DPT_L-384 and averaging over the datasets. So, if $\epsilon_d$ is the zero-shot error for dataset $d$, then the $\color{green}{\textsf{improvement}}$ is given by $100(1-(1/6)\sum_d\epsilon_d/\epsilon_{d,\rm{DPT_{L-384}}})$%.

Note that the improvements of 10% for MiDaS v2.0 → v2.1 and 21% for MiDaS v2.1 → v3.0 are not visible from the improvement column (Imp.) in the table but would require an evaluation with respect to MiDaS v2.1 Large₃₈₄ and v2.0 Large₃₈₄ respectively instead of v3.0 DPT_L-384.

Depth map comparison

Zoom in for better visibility

Speed on Camera Feed

Test configuration

Windows 10
11th Gen Intel Core i7-1185G7 3.00GHz
16GB RAM
Camera resolution 640x480
openvino_midas_v21_small_256

Speed: 22 FPS

Applications

MiDaS is used in the following other projects from Intel Labs:

ZoeDepth (code available here): MiDaS computes the relative depth map given an image. For metric depth estimation, ZoeDepth can be used, which combines MiDaS with a metric depth binning module appended to the decoder.
LDM3D (Hugging Face model available here): LDM3D is an extension of vanilla stable diffusion designed to generate joint image and depth data from a text prompt. The depth maps used for supervision when training LDM3D have been computed using MiDaS.

Changelog

[Dec 2022] Released MiDaS v3.1:
- New models based on 5 different types of transformers (BEiT, Swin2, Swin, Next-ViT, LeViT)
- Training datasets extended from 10 to 12, including also KITTI and NYU Depth V2 using BTS split
- Best model, BEiT_{Large 512}, with resolution 512x512, is on average about 28% more accurate than MiDaS v3.0
- Integrated live depth estimation from camera feed
[Sep 2021] Integrated to Huggingface Spaces with Gradio. See Gradio Web Demo.
[Apr 2021] Released MiDaS v3.0:
- New models based on Dense Prediction Transformers are on average 21% more accurate than MiDaS v2.1
- Additional models can be found here
[Nov 2020] Released MiDaS v2.1:
- New model that was trained on 10 datasets and is on average about 10% more accurate than MiDaS v2.0
- New light-weight model that achieves real-time performance on mobile platforms.
- Sample applications for iOS and Android
- ROS package for easy deployment on robots
[Jul 2020] Added TensorFlow and ONNX code. Added online demo.
[Dec 2019] Released new version of MiDaS - the new model is significantly more accurate and robust
[Jul 2019] Initial release of MiDaS (Link)

Citation

Please cite our paper if you use this code or any of the models:

@ARTICLE {Ranftl2022,
    author  = "Ren\'{e} Ranftl and Katrin Lasinger and David Hafner and Konrad Schindler and Vladlen Koltun",
    title   = "Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross-Dataset Transfer",
    journal = "IEEE Transactions on Pattern Analysis and Machine Intelligence",
    year    = "2022",
    volume  = "44",
    number  = "3"
}

If you use a DPT-based model, please also cite:

@article{Ranftl2021,
    author    = {Ren\'{e} Ranftl and Alexey Bochkovskiy and Vladlen Koltun},
    title     = {Vision Transformers for Dense Prediction},
    journal   = {ICCV},
    year      = {2021},
}

Please cite the technical report for MiDaS 3.1 models:

@article{birkl2023midas,
      title={MiDaS v3.1 -- A Model Zoo for Robust Monocular Relative Depth Estimation},
      author={Reiner Birkl and Diana Wofk and Matthias M{\"u}ller},
      journal={arXiv preprint arXiv:2307.14460},
      year={2023}
}

For ZoeDepth, please use

@article{bhat2023zoedepth,
  title={Zoedepth: Zero-shot transfer by combining relative and metric depth},
  author={Bhat, Shariq Farooq and Birkl, Reiner and Wofk, Diana and Wonka, Peter and M{\"u}ller, Matthias},
  journal={arXiv preprint arXiv:2302.12288},
  year={2023}
}

and for LDM3D

@article{stan2023ldm3d,
  title={LDM3D: Latent Diffusion Model for 3D},
  author={Stan, Gabriela Ben Melech and Wofk, Diana and Fox, Scottie and Redden, Alex and Saxton, Will and Yu, Jean and Aflalo, Estelle and Tseng, Shao-Yen and Nonato, Fabio and Muller, Matthias and others},
  journal={arXiv preprint arXiv:2305.10853},
  year={2023}
}

Acknowledgements

Our work builds on and uses code from timm and Next-ViT. We'd like to thank the authors for making these libraries available.

License

MIT License