ITSA (CVPR 2022)

This is the official code of the work ITSA: An Information-Theoretic Approach to Automatic Shortcut Avoidance and Domain Generalization in Stereo Matching Networks, CVPR 2022,
WeiQin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, Ali Bab-Hadiashar, David Suter. [Arxiv]

How to use:

Environment setup

• Python 3.8
• PyTorch 1.9.0

📦 Dataset

• KITTI 2015
• Scene Flow
• Middlebury
• ETH3D

🕓 Training

Run bash script/train.sh for training.

Arguments

• itsa: Path to the pre-trained weight file that you wish to load to the model during training
• model: Select the stereo matching networks from one of the following: [PSMNet, GwcNet, CFNet].
• maxdisp: Range of disparity. Default = 192.
• epochs: Total number of training epochs.
• lambd: Hyperparameter for our Fisher loss.
• eps: Shortcut-perturbation augmentation strength.

📝 Inference

Run bash eval_kitti.sh for inference. Predicted disparity maps for the chosen dataset will be saved to the directory set in the argument savepath.

Arguments

• loadmodel: Path to the pre-trained weight file that you wish to load to the model during training. If directory, all checkpoints will be used for evaluation in loops.
• savepath: Path to save the estimated disparity maps. If None, nothing will be saved.
• datapath: Path to the selected data. Make sure you have downloaded and extracted the required dataset to a specific location in your local machine.
• model: Select the stereo matching networks from one of the following: [PSMNet, GwcNet, CFNet].
• Select the domain you want to evaluate accordingly [kitti15, kitti12, midFull, midHalf, midQuar, eth] (see below)

Example

To run inference using KITTI 2015 dataset and PSMNet, run the following command:

python infer.py --model PSMNet \
                --savepath {path_to_save_images} \
                --loadmodel {path_to_pretrained_weights}\
                --kitti15 {change this to the desired dataset}

Make sure the directory to the datasets are correct.

Pretrained Weights [Google Drive]

• PSMNet
• GwcNet
• CFNet

Citation

If you find this code useful in your research, please cite:

@InProceedings{Chuah_2022_CVPR,
    author    = {Chuah, WeiQin and Tennakoon, Ruwan and Hoseinnezhad, Reza and Bab-Hadiashar, Alireza and Suter, David},
    title     = {ITSA: An Information-Theoretic Approach to Automatic Shortcut Avoidance and Domain Generalization in Stereo Matching Networks},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {13022-13032}
}

Acknowledgement

The codes in this work heavily relies on codes by PSMNet, GwcNet and CFNet. We thank the original authors for their awesome repos.