OptimizeMVS
OptimizeMVS copied to clipboard
[CVPR 2019] Conditional Single-view Shape Generation for Multi-view Stereo Reconstruction
OptimizeMVS
Created by Yi Wei, Shaohui Liu and Wang Zhao from Tsinghua University.
Introduction
This repository contains source code for Conditional Single-view Shape Generation for Multi-view Stereo Reconstruction in tensorflow.
Installation
The code has been tested with Python 2.7, tensorflow 1.3.0 on Ubuntu 16.04.
1. Clone code
git clone https://github.com/weiyithu/OptimizeMVS.git
2. Install packages
Python virtual environment is recommended.
cd OptimizeMVS
virtualenv env
source ./env/bin/activate
pip install -r requirements.txt
You need to compile the external libraries:
sh compile.sh
Usage
1. One-button setup
sh init_data.sh
2. Two-stage training
sh train.sh
Default training options are stored in the config
folder.
3. Evaluation
You can download our pretrained model: single-category, multi-category and move it to a folder named demo
for evaluation.
sh download_trained_model.sh
To evaluate your own model, set the 'load' option in test.sh
as the path to your model.
sh test.sh
The results might be slightly better than reported in the paper.
Acknowledgements
Part of the external operators are borrowed from latent_3d_points and PointNet++. The multi-view images were rendered from ShapeNetCore with the preprocessing scripts in mvcSnP and the point cloud data was from latent_3d_points. We sincerely thank the authors for their kind help.
This work was supported in part by the National Natural Science Foundation of China under Grant U1813218, Grant 61822603, Grant U1713214, Grant 61672306, and Grant 61572271.
Citation
If you find this work useful in your research, please consider citing:
@inproceedings{wei2019conditional,
author = {Wei, Yi and Liu, Shaohui and Zhao, Wang and Lu, Jiwen and Zhou, Jie},
title = {Conditional Single-view Shape Generation for Multi-view Stereo Reconstruction},
booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2019}
}
The first three authors share equal contributions.