RayMVSNet: Learning Ray-based 1D Implicit Fields for Accurate Multi-View Stereo

Junhua Xi* Yifei Shi* Yijie Wang Yulan Guo Kai Xu†

National University of Defense Technology

• CUDA 11.2
• Python 3.8.5
• torch 1.7.1+cu110
• pip install -r requirements.txt

• Download the preprocessed DTU data and unzip it to data/dtu.

./dtu  
      ├── Rectified                 
      │   ├── scan1_train       
      │   ├── scan2_train       
      │   └── ...                
      ├── Cameras
      │   ├── pair.txt   
      │   ├── train   
      │       ├── 00000000_cam.txt   
      │       ├── 00000001_cam.txt   
      │       └── ...  
      └── Depths         
          ├── scan1_train   
          ├── scan2_train    
          └── ... 

• Train the network
• python train.py

• python test.py

You can test with the pretrained model: ./model.ckpt

Testing process generates per-view depth map. We need to apply depth fusion fusion.py to get the complete point cloud. Please refer to MVSNet for more details.

• python fusion.py

• Download the offical evaluation tool from DTU benchmark
• We provide our pre-computed point clouds for your convenience