Self-supervised Learning of Depth Inference for Multi-view Stereo (CVPR 2021)

This repository extend the original CVP-MVSNet with unsupervised training and self-training.

Please refer to the original CVP-MVSNet for basic usage.

Pre-requisites for unsupervised initialization and self-training:

• open3d
• trimesh
• pyrender

Coming soon...

Code and scripts for self-training can be found in the CVP-MVSNet/self-training folder.

• self-training.sh: Shell script to run following code and generate pseudo depth from a given checkpoint.
• pseudo_fusion.py: Pseudo label filtering and Multi-view pseudo label fusion.
• surface_reconstruction.py: Generate pseudo mesh by Screened Poisson Surface Reconstruction.
• pseudo_render.py: Render pseudo mesh into new pseudo depth maps.

To generate pseudo depth given a base checkpoint:

sh self-training.sh

Pseudo depth will be generated in self-training/outputs_self_training_itr$ITR/pseudo_depth_128/ folder.

The generated pseudo depth can be linked into dataset/dtu-train-128/ and replace the existing Depths folder using either mv or ln -s.

Next iteration of self-training can be start by train.sh.

If you find this project useful for your research, please cite:

@inproceedings{Yang_2021_CVPR,
  author = {Yang, Jiayu and Alvarez, Jose M. and Liu, Miaomiao},
  title={Self-supervised Learning of Depth Inference for Multi-view Stereo},
  booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}

@InProceedings{Yang_2020_CVPR,
    author = {Yang, Jiayu and Mao, Wei and Alvarez, Jose M. and Liu, Miaomiao},
    title = {Cost Volume Pyramid Based Depth Inference for Multi-View Stereo},
    booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year = {2020}
}