Convolutional Spatial Propagation Network

By Xinjing Cheng, Peng Wang, Chenye Guan and Ruigang Yang

Contents

1. Introduction
2. Citation

Introduction

This repo contains the CSPN models trained for depth completion and stereo depth estimation, as as described in the paper “Depth Estimation via Affinity Learned with Convolutional Spatial Propagation Network“ and “
Learning Depth with Convolutional Spatial Propagation Network“

We provide two implementations here. First, we implement a naive 2D CSPN by Pytorch which combine ‘offsite’ and ‘conv’ operations. Then, we provide a more efficient implementation of 2D/3D CSPN by CUDA and integrated them to the PaddlePaddle, which is more stable and easy to use.

Citation

If you use this method in your research, please cite:

@inproceedings{cheng2018depth,
  title={Depth estimation via affinity learned with convolutional spatial propagation network},
  author={Cheng, Xinjing and Wang, Peng and Yang, Ruigang},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  pages={103--119},
  year={2018}
}

@article{cheng2019learning,
  title={Learning depth with convolutional spatial propagation network},
  author={Cheng, Xinjing and Wang, Peng and Yang, Ruigang},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  year={2019},
  publisher={IEEE}
}