Project Page | Paper | Video
This repo is the official implementation of Digging Into Normal Incorporated Stereo Matching, ACM MM2022
In this paper, we propose a normal incorporated joint learning framework consisting of two specific modules named non-local disparity propagation(NDP) and affinity-aware residual learning(ARL). The estimated normal map is first utilized for calculating a non-local affinity matrix and a non-local offset to perform spatial propagation at the disparity level. To enhance geometric consistency, especially in low-texture regions, the estimated normal map is then leveraged to calculate a local affinity matrix, providing the residual learning with information about where the correction should refer and thus improving the residual learning efficiency.
- Python 3.9, PyTorch >= 1.8.0
- CUDA ToolKit for DCN-V2 Compile
- TODO
- TODO
@inproceedings{liu2022digging,
title={Digging Into Normal Incorporated Stereo Matching},
author={Liu, Zihua and Zhang, Songyan and Wang, Zhicheng and Okutomi, Masatoshi},
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
pages={6050--6060},
year={2022}
}