Awesome

PVA-MVSNet

About

This official repository is established for Pyramid Multi-view Stereo Net with Self-adaptive View Aggregation (ECCV2020) paper

How to Use

Requirements

python 3.6
Pytorch >= 1.0.0
CUDA >= 9.0

Install

./conda_install.sh

Training

Download the preprocessed DTU training data (also available at Baiduyun, code: s2v2), and upzip it as the MVS_TRANING folder (orrowed from MVSNet(https://raw.githubusercontent.com/YoYo000/MVSNet)).
Set dtu_data_root to your MVS_TRAINING path in env.sh Create a log folder and a model folder in wherever you like to save the training outputs. Set the log_dir and save_dir in train.sh correspondingly.
Train VA-MVSNet (GTX1080Ti): ./train.sh

Testing

Download the test data for scan9 and unzip it as the TEST_DATA_FOLDER folder, which should contain one cams folder, one images folder and one pair.txt file.
Download the pre-trained VA-MVSNet models and upzip the file as MODEL_FOLDER
In eval_pyramid.sh, set MODEL_FOLDER to ckpt and model_ckpt_index to checkpoint_list.
Run ./eval_pyramid.sh.

MMP and Filter&Fusion

We utilize depthfusion_pytorch.py script for Fusion (from MVSNet-pytorch).
Set use_mmp as True to use Multi-metric Pyramid Depth Aggregation in tools/postprocess.sh.
Enter to ./tools directory, then run ./postprocess.sh to generate final point cloud.

Reproduce Benchmark results

Results on DTU

	Acc.	Comp.	Overall.
MVSNet(D=256)	0.396	0.527	0.462
PVAMVSNet(D=192)	0.379	0.336	0.357

PVA-MVSNet point cloud results with full post-processing are also provided: DTU evaluation point clouds with extracting code zau7.

Results on Tanks and Temples

Mean	Family	Francis	Horse	Lighthouse	M60	Panther	Playground	Train
54.46	69.36	46.80	46.01	55.74	57.23	54.75	56.70	49.06

Please ref to leaderboard.

Citation

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

@inproceedings{yi2020PVAMVSNET,
  title={Pyramid multi-view stereo net with self-adaptive view aggregation},
  author={Yi, Hongwei and Wei, Zizhuang and Ding, Mingyu and Zhang, Runze and Chen, Yisong and Wang, Guoping and Tai, Yu-Wing},
  booktitle={ECCV},
  year={2020}
}

Acknowledgement

Thanks Xiaoyang Guo for his contribution to re-implementation of MVSNet-pytorch. Thanks Yao Yao for his previous works MVSNet/R-MVSNet.