Spatio-Focal Bidirectional Disparity Estimation from a Dual-Pixel Image

Project page | Paper | Supplemental

Donggun Kim, Hyeonjoong Jang, Inchul Kim, Min H. Kim

Demo code and captured dual-pixel blur kernel for the paper "Spatio-Focal Bidirectional Disparity Estimation from a Dual-Pixel Image", presented at CVPR 2023.

@InProceedings{Kim:2023:Dualpixel,
    title = {Spatio-Focal Bidirectional Disparity Estimation from a Dual-Pixel Image},
    author = {Kim, Donggun and Jang, Hyeonjoong and Kim, Inchul and Kim, Min H.},
    booktitle = {{IEEE} Conference on Computer Vision and Pattern Recognition ({CVPR})},
    year = {2023}
}

Contents

0. Prepare

1. Environment setup

2. Run

3. Dual-Pixel Blur Kernel

Troubleshooting & Misc. Info

0. Prepare

Dataset

For the dataset, we assume the following structure and root data directory path are used for the following steps.

/where/your/data/dir/root
├── Abuolaim_ECCV_2020 (1.3GB, canon test data only)
│   └── canon
│       ├── test_l
│       │   └── source
│       └── test_r
│           └── source
├── Punnappurath_ICCP_2020 (600 MB)
│   ├── 001_B.jpg
│   ├── 001_D.TIF
│   ├── 001_L.jpg
│   ├── 001_R.jpg
│   ├── 002_B.jpg
│   └── ...
└── Xin_ICCV_2021 (185 MB)
    ├── 001_left.png
    ├── 001_right.png
    ├── 002_left.png
    ├── ...
    └── gt_defocus_map

If you don't have the dataset, the download link is in the Acknowledgement section. If you want to use your own data, see Using own data section.

Checkpoint

The checkpoint is already included in this git repository (ckpt/checkpoint.pt, 43MB).

1. Environment setup

Make sure that your hardware can run PyTorch 1.11.0. We provide 3 options. We highly recommend option 1 (Docker with CUDA GPU support) to prevent conflict with your local environment.

(a) Docker CUDA support (NVIDIA Container Toolkit): Available for Linux / Windows 11 WSL. In case of Windows 10 WSL, you need 21H2 update [Linux reference, Windows reference].

Option 1. Docker + CUDA (Recommended)

Make sure you properly installed Docker CUDA support. The following command builds a docker image with env/Dockerfile. This may take some time.

docker build -t sfbd:v0 ./env

Set the path as shell variable for starting the docker container. See the data directory structure in the Prepare dataset section.

Note that $(pwd) automatically get current directory and you have to put your dataset directory to where/your/data/dir/root.

THIS_REPOSITORY_PATH="$(pwd)"
DATA_DIR_ROOT="/where/your/data/dir/root"

Start the docker container. You may modify some options (e.g. --name for container name (2nd line), --gpu for gpu index (3rd line), additional -v option for mount volume (4th line)).

docker run -ti \
--name sfbd \
--gpus all \
-v "$THIS_REPOSITORY_PATH":/workspace/code \
-v "$DATA_DIR_ROOT":/workspace/data \
sfbd:v0

Now you are inside the docker container. Inside the docker container, your code exists in /workspace/code and data exists in /workspace/data.

The following command installs the remaining dependencies. Everything is prepared in the base environment of docker. Therefore, this does NOT create the new conda environment.

cd /workspace/code
chmod +x env/install.sh && ./env/install.sh

Option 2. Non-Docker + CUDA

(NB: This is not tested) If you are not going to use Docker, ensure the following environment matches.

• Pytorch 1.11.0 compatible, CUDA Toolkit version 11.3 and CuDNN 8
• Anaconda installed
• (Windows) MinTTY (usually installed with Git Bash) with conda enabled, for running the following shell script

And run the following commands. This will create the NEW conda environment named sfbd.

chmod +x env/install_conda_cuda.sh && ./env/install_conda_cuda.sh

Option 3. CPU-only

(NB: This is extremely slow) In case of CUDA-enabled GPU is not available, you can use the CPU-only version. In this case, you need an environment similar to Option 2, but without CUDA support.

And run the following commands. This will create the NEW conda environment named sfbd.

chmod +x env/install_conda_cpu.sh && ./env/install_conda_cpu.sh

2. Run

First ensure that you are in the right docker container (for option 1) and conda environment (base for option 1, sfbd for option 2 and 3).

Run demo one by one using demo.py. Following is one example of a running command.

python demo.py -data_name "CanonDPDepth" -save_disp -save_invdepth -l "../data/Punnappurath_ICCP_2020/*_L.jpg" -r "../data/Punnappurath_ICCP_2020/*_R.jpg" -gt "../data/Punnappurath_ICCP_2020/*_D.TIF"

See demo.py for details and input arguments and demo_all.sh for examples.

Using own data

you can use the class DPDataset (i.e., -data_name DPDataset) or create your own class inheriting the DPDataset class. See dpdataset.py for more detail.

3. Dual-Pixel Blur Kernel

We also provide blur kernel experiment result in kernel directory. See kernel/README.md for detail.

Troubleshooting & Misc. Info

Install script failure in Windows Docker

Maybe you have error message: bash: ./install.sh: /bin/bash^M: bad interpreter: No such file or directory. This problem is caused since you are using code from Windows host OS mounted to Linux (Docker). When you are cloning the git repository, if you didn't configure the git properly, it will automatically convert Unix line break LF to CR LF. To solve this, you should convert CR LF to the LF. Simplest way to do this is run the following command inside the Docker, or use external text editor (e.g. VSCode) to convert it.

sed -i 's/\r//' env/*.sh

Timeout during downloading torch via pip

You can solve this by increasing the timeout. In env/install_conda*.sh, increase the number (in seconds) with --timeout. The default is set to 180. If it is not sufficient, increase as much as possible.

... UserWarning: torch.meshgrid ...

This is just a warning (i.e., no effects on running). You can ignore this if you use the correct version of pytorch (1.11.0). If you want to remove this warning, add indexing='ij' option to all torch.meshgrid function in python files under core/.

Metrics from this repository are better than the main paper Table 2

Metrics from this repository (AI1=0.0383, AI2=0.0667, 1-|spcc|=0.2527) are slightly better than the values in the paper. The difference comes from the input image quality. We use pre-resized images with compression (jpg) in the paper, while this repository performs only resizing from the original image.

License

Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H. Kim have developed this software and related documentation (the "Software"); confidential use in source form of the Software, without modification, is permitted provided that the following conditions are met:

Neither the name of the copyright holder nor the names of any contributors may be used to endorse or promote products derived from the Software without specific prior written permission.

The use of the software is for Non-Commercial Purposes only. As used in this Agreement, “Non-Commercial Purpose” means for the purpose of education or research in a non-commercial organization only. “Non-Commercial Purpose” excludes, without limitation, any use of the Software for, as part of, or in any way in connection with a product (including software) or service which is sold, offered for sale, licensed, leased, published, loaned or rented. If you require a license for a use excluded by this agreement, please email minhkim@kaist.ac.kr.

Warranty: KAIST-VCLAB MAKES NO REPRESENTATIONS OR WARRANTIES ABOUT THE SUITABILITY OF THE SOFTWARE, EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, OR NON-INFRINGEMENT. KAIST-VCLAB SHALL NOT BE LIABLE FOR ANY DAMAGES SUFFERED BY LICENSEE AS A RESULT OF USING, MODIFYING OR DISTRIBUTING THIS SOFTWARE OR ITS DERIVATIVES.

Please refer to license for more details.

Acknowledgement

This project uses the following open-source projects and data. Please consider citing them if you use related functionalities.

• RAFT-stereo (Lipson et al., 3DV 2021)
• Garg et al., ICCV 2019

Dataset:

• Punnappurath et al., ICCP 2020 (Download link is in the "Dataset" section)
• Abuolaim et al., ECCV 2020 (Download link is in the "Dataset" section, "All images used for training/testing")
• Xin et al., ICCV 2021 (Data exists in the "DP_data_pixel_4" directory)