Home

Awesome

Fast-ACVNet

Our significant extension version of ACV, named Fast-ACV, is available at Paper, Code

MethodScene Flow <br> (EPE)KITTI 2012 <br> (3-all)KITTI 2015 <br> (D1-all)Runtime (ms)
Fast-ACVNet+0.591.85 %2.01 %45
HITNet-1.89 %1.98 %54
CoEx0.691.93 %2.13 %33
BGNet+-2.03 %2.19 %35
AANet0.872.42 %2.55 %62
DeepPrunerFast0.97-2.59 %50

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

ACVNet (CVPR 2022)

This is the implementation of the paper: Attention Concatenation Volume for Accurate and Efficient Stereo Matching, CVPR 2022, Gangwei Xu, Junda Cheng, Peng Guo, Xin Yang

Introduction

An informative and concise cost volume representation is vital for stereo matching of high accuracy and efficiency. In this paper, we present a novel cost volume construction method which generates attention weights from correlation clues to suppress redundant information and enhance matching-related information in the concatenation volume. To generate reliable attention weights, we propose multi-level adaptive patch matching to improve the distinctiveness of the matching cost at different disparities even for textureless regions.

image

How to use

Environment

Install

Create a virtual environment and activate it.

conda create -n acvnet python=3.8
conda activate acvnet

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

Data Preparation

Download Scene Flow Datasets, KITTI 2012, KITTI 2015

Train

Use the following command to train ACVNet on Scene Flow

Firstly, train attention weights generation network for 64 epochs,

python main.py --attention_weights_only True

Secondly, freeze attention weights generation network parameters, train the remaining network for another 64 epochs,

python main.py --freeze_attention_weights True

Finally, train the complete network for 64 epochs,

python main.py

Use the following command to train ACVNet on KITTI (using pretrained model on Scene Flow)

python main_kitti.py

Test

python test_sceneflow.py

Pretrained Model

Scene Flow

Results on KITTI 2015 leaderboard

Leaderboard Link

MethodD1-bg (All)D1-fg (All)D1-all (All)Runtime (s)
ACVNet1.37 %3.07 %1.65 %0.20
LEAStereo1.40 %2.91 %1.65 %0.30
GwcNet1.74 %3.93 %2.11 %0.32
PSMNet1.86 %4.62 %2.32 %0.41

Qualitative results on Scene Flow Datasets, KITTI 2012 and KITTI 2015

The left column is left image, and the right column is results of our ACVNet.

image

Citation

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

@inproceedings{xu2022attention,
  title={Attention Concatenation Volume for Accurate and Efficient Stereo Matching},
  author={Xu, Gangwei and Cheng, Junda and Guo, Peng and Yang, Xin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={12981--12990},
  year={2022}
}

@article{xu2023accurate,
  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={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2023},
  publisher={IEEE}
}

Acknowledgements

Thanks to Xiaoyang Guo for opening source of his excellent work GwcNet. Our work is inspired by this work and part of codes are migrated from GwcNet.