ISSGA-Stereo

Pytorch implementation of the paper "Inter-Scale Similarity Guided Cost Aggregation for Stereo Matching". This repo is keeping update.

I've been involved in a car accident recently. As a result, I have sustained a fracture in my right hand. This unfortunate incident is expected to impact the timeline for the upcoming code updates. 😭

For the core code, please refer to ./models./issga.py

Training

python train.py --maxdisp 384 --batchsize 6 --database data --savemodel ./checkpoints  --epochs 30 

Evaluation

Sceneflow

CUDA_VISIBLE_DEVICES=0 python test_sceneflow_raw.py --maxdisp 192 --database ./data --loadmodel  "./checkpoints/sceneflow.tar"

Environment

• NVIDIA RTX 3090
• Python 3.8
• Pytorch 1.19

Dependencies

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -c nvidia
pip install opencv-python
pip install scikit-image
pip install tensorboardX
pip install matplotlib 

Required Data

To evaluate/train ISSGA-Stereo, you will need to download the required datasets.

• Scene Flow
• KITTI
• Middlebury
• ETH3D

├── /data
    ├── sceneflow
        ├── frames_finalpass
        ├── disparity
    ├── KITTI
        ├── KITTI_2012
            ├── training
            ├── testing
            ├── vkitti
        ├── KITTI_2015
            ├── training
            ├── testing
            ├── vkitti
    ├── Middlebury
        ├── trainingH
        ├── trainingH_GT
        ├── trainingQ
        ├── trainingQ_GT
    ├── ETH3D
        ├── two_view_training
        ├── two_view_training_gt

Citation

If you find our work useful in your research, please consider citing our paper:

@inproceedings{issga-stereo,
  title={Inter-Scale Similarity Guided Cost Aggregation for
Stereo Matching},
  author={Pengxiang Li, Chengtang Yao, Yunde Jia, and Yuwei Wu},
  year={2023}
}

Acknowledgements

This project is heavily based on HSM-Net and CF-Net, we thank the original authors for their excellent work.