MVSNet: Depth Inference for Unstructured Multi-view Stereo. Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, Long Quan. ECCV 2018. MVSNet is a deep learning architecture for depth map inference from unstructured multi-view images.

This is an unofficial Pytorch implementation of MVSNet

• python 3.6 (Anaconda)
• pytorch 1.0.1

• Download the preprocessed DTU training data (Fixed training cameras, from Original MVSNet), and upzip it as the MVS_TRANING folder
• in train.sh, set MVS_TRAINING as your training data path
• create a logdir called checkpoints
• Train MVSNet: ./train.sh

• Download the preprocessed test data DTU testing data (from Original MVSNet) and unzip it as the DTU_TESTING folder, which should contain one cams folder, one images folder and one pair.txt file.
• in test.sh, set DTU_TESTING as your testing data path and CKPT_FILE as your checkpoint file. You can also download my pretrained model.
• Test MVSNet: ./test.sh

in eval.py, I implemented a simple version of depth map fusion. Welcome contributions to improve the code.

Due to the memory limit, we only train the model with D=192, the fusion code is also different from the original repo.