MA-Stereo: Real-Time Stereo Matching via Multi-scale Attention Fusion and Spatial Error-aware Refinement

Wei Gao, Yongjie Cai, Youssef Akoudad, Yang Yang, and Jie Chen

How to use

Data Preparation

• Scene Flow Datasets
• KITTI 2012
• KITTI 2015
• Middlebury
• ETH3D

Train

Use the following command to train MA-Stereo on Scene Flow. First training,

python train_sceneflow.py --logdir ./checkpoints/sceneflow/first/

Second training,

python train_sceneflow_2.py --logdir ./checkpoints/sceneflow/second/ --loadckpt ./checkpoints/sceneflow/first/checkpoint_000063.ckpt

Use the following command to train CGI-Stereo on KITTI (using pretrained model on Scene Flow),

python train_kitti.py --logdir ./checkpoints/kitti/ --loadckpt ./checkpoints/sceneflow/second/checkpoint_000063.ckpt

Evaluation on Scene Flow and KITTI

Pretrained Model

• Google Drive

Generate disparity images of KITTI test set,

python save_disp.py

Acknowledgement

Special thanks to CGI-Stereo for providing the code base for this work.

CGI-Stereo