Accurate and Efficient Stereo Matching via Attention Concatenation Volume

Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, Xin Yang

Paper

Fast-ACVNet.

Demo on KITTI raw data

A demo result on our RTX 3090 (Ubuntu 20.04).

How to use

Environment

• Python 3.8
• Pytorch 1.10

Install

Create a virtual environment and activate it.

conda create -n fast_acv python=3.8
conda activate fast_acv

Dependencies

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch -c nvidia
pip install opencv-python
pip install scikit-image
pip install tensorboard
pip install matplotlib 
pip install tqdm
pip install timm==0.5.4

Data Preparation

Download Scene Flow Datasets, KITTI 2012, KITTI 2015

Train

Use the following command to train Fast-ACVNet+ or Fast-ACVNet on Scene Flow

Firstly, train attention weights generation network for 24 epochs,

python main_sceneflow.py --attention_weights_only True --logdir ./checkpoints/sceneflow/attention

Secondly, train complete network for another 24 epochs,

python main_sceneflow.py --loadckpt ./checkpoints/sceneflow/attention/checkpoint_000023.ckpt --logdir ./checkpoints/sceneflow/complete

Use the following command to train Fast-ACVNet+ or Fast-ACVNet on KITTI (using pretrained model on Scene Flow),

python main_kitti.py --loadckpt ./checkpoints/sceneflow/complete/checkpoint_000023.ckpt --logdir ./checkpoints/kitti

Submitted to KITTI benchmarks

python save_disp.py

Evaluation on Scene Flow and KITTI

Our Fast-ACVNet+ achieves comparable accuracy with HITNet on KITTI 2012 and KITTI 2015

Pretrained Model

Fast-ACVNet

Fast-ACVNet+

Qualitative results on Scene Flow.

Qualitative results on KITTI.

Generalization performance on the Middlebury 2014 dataset. All the comparison methods are only trained on Scene Flow without data augmentation.

Citation

If you find this project helpful in your research, welcome to cite the paper.

@article{xu2022accurate,
  title={Accurate and Efficient Stereo Matching via Attention Concatenation Volume},
  author={Xu, Gangwei and Wang, Yun and Cheng, Junda and Tang, Jinhui and Yang, Xin},
  journal={arXiv preprint arXiv:2209.12699},
  year={2022}
}

Acknowledgements

Thanks to Antyanta Bangunharcana for opening source of his excellent work Correlate-and-Excite. Thanks to Xiaoyang Guo for opening source of his excellent work GwcNet.