Skip to content


Repository files navigation

PatchmatchNet (CVPR2021 Oral)

official source code of paper 'PatchmatchNet: Learned Multi-View Patchmatch Stereo'


  • 13.12.2021: New unified format for training and evaluation datasets, support for arbitrary image sizes and multi-camera setups, and new names for script parameters.
  • 27.09.2021: The code now allows for Torchscript export and includes a pre-trained TorchScript module.


PatchmatchNet is a novel cascade formulation of learning-based Patchmatch which aims at decreasing memory consumption and computation time for high-resolution multi-view stereo. If you find this project useful for your research, please cite:

      title={PatchmatchNet: Learned Multi-View Patchmatch Stereo}, 
      author={Fangjinhua Wang and Silvano Galliani and Christoph Vogel and Pablo Speciale and Marc Pollefeys},



  • python 3.8
  • CUDA >= 10.1
pip install -r requirements.txt

Reproducing Results

├──scan1 (scene_name1)
├──scan2 (scene_name2) 
      ├── images                 
      │   ├── 00000000.jpg       
      │   ├── 00000001.jpg       
      │   └── ...                
      ├── cams                   
      │   ├── 00000000_cam.txt   
      │   ├── 00000001_cam.txt   
      │   └── ...                
      └── pair.txt  


  • The subfolders for Tanks & Temples and ETH3D will not be named scanN but the lists included under ./lists/eth3d and ./lists/tanks will have the correct naming conventions.
  • If the folders for images and cameras, and the pair file don't follow the standard naming conventions you can modify the settings of MVSDataset in datasets/ to specify the custom image_folder, cam_folder, and pair_path
  • The MVSDataset is configured by default for JPEG images. If you're using a different format (e.g., PNG) you can change the image_extension parameter of MVSDataset accordingly.

Camera file cam.txt stores the camera parameters, which includes extrinsic, intrinsic, minimum depth and maximum depth:

E00 E01 E02 E03
E10 E11 E12 E13
E20 E21 E22 E23
E30 E31 E32 E33

K00 K01 K02
K10 K11 K12
K20 K21 K22


pair.txt stores the view selection result. For each reference image, N (10 or more) best source views are stored in the file:

IMAGE_ID0                       # index of reference image 0 
10 ID0 SCORE0 ID1 SCORE1 ...    # 10 best source images for reference image 0 
IMAGE_ID1                       # index of reference image 1
10 ID0 SCORE0 ID1 SCORE1 ...    # 10 best source images for reference image 1 
  • In, set DTU_TESTING, ETH3D_TESTING or TANK_TESTING as the root directory of corresponding dataset and uncomment the evaluation command for corresponding dataset (default is to evaluate on DTU's evaluation set). If you want to change the output location (default is same as input one), modify the --output_folder parameter. For Tanks the --scan_list can be intermediate or advanced and for ETH3D it can be test or train.
  • CKPT_FILE is the checkpoint file (our pretrained model is ./checkpoints/params_000007.ckpt), change it if you want to use your own model. If you want to use the model from the TorchScript module instead, you can specify the checkpoint file as ./checkpoints/ and set the option --input_type module.
  • Test on GPU by running sh The code includes depth map estimation and depth fusion. The outputs are the point clouds in ply format.
  • For quantitative evaluation on DTU dataset, download SampleSet and Points. Unzip them and place Points folder in SampleSet/MVS Data/. The structure looks like:
├──MVS Data

In evaluations/dtu/BaseEvalMain_web.m, set dataPath as path to SampleSet/MVS Data/, plyPath as directory that stores the reconstructed point clouds and resultsPath as directory to store the evaluation results. Then run evaluations/dtu/BaseEvalMain_web.m in matlab.

The results look like:

Acc. (mm) Comp. (mm) Overall (mm)
0.427 0.277 0.352
  • For detailed quantitative results on Tanks & Temples and ETH3D, please check the leaderboards (Tanks & Temples, ETH3D)

Evaluation on Custom Dataset

  • For evaluation, we support preparing the custom dataset from COLMAP's results. The script (modified based on the script from MVSNet) converts COLMAP's sparse reconstruction results into the same format as the datasets that we provide. After reconstruction, COLMAP will generate a folder COLMAP/dense/, which contains COLMAP/dense/images/ and COLMAP/dense/sparse. Then you need to run like this:
python --input_folder COLMAP/dense/ 
  • The default output location is the same as the input one. If you want to change that, set the --output_folder parameter.
  • The default behavior of the converter will find all possible related images for each source image. If you want to constrain the max number of related images set the --num_src_images parameter.
  • In, set CUSTOM_TESTING as the root directory of the dataset, set --output_folder as the directory to store the reconstructed point clouds (default is same as input directory), set --image_max_dim to an appropriate size (this is determined by the available GPU memory and the desired processing speed) or use the native size by removing the parameter, and uncomment the evaluation command. Test on GPU by running sh


Download pre-processed DTU's training set. The dataset is already organized as follows:

├── Cameras_1
│    ├── train
│    │    ├── 00000000_cam.txt
│    │    ├── 00000000_cam.txt
│    │    └── ...
│    └── pair.txt
├── Depths_raw
│    ├── scan1
│    │    ├── depth_map_0000.pfm
│    │    ├── depth_visual_0000.png
│    │    ├── depth_map_0001.pfm
│    │    ├── depth_visual_0001.png
│    │    └── ...
│    ├── scan2
│    └── ...
└── Rectified
     ├── scan1_train
     │    ├── rect_001_0_r5000.png
     │    ├── rect_001_1_r5000.png
     │    ├── ...
     │    ├── rect_001_6_r5000.png
     │    ├── rect_002_0_r5000.png
     │    ├── rect_002_1_r5000.png
     │    ├── ...
     │    ├── rect_002_6_r5000.png
     │    └── ...
     ├── scan2_train
     └── ...

To use this dataset directly look into the Legacy Training section below. For the current version of training the dataset needs to be converted to a format compatible with MVSDataset in ./datasets/ using the script as follows:

python --input_folder <original_dataset> --output_folder <converted_dataset> --scan_list ./lists/dtu/all.txt

The converted dataset will now be in a format similar to the evaluation datasets:

├── scan1 (scene_name1)
├── scan2 (scene_name2) 
│     ├── cams (camera parameters)
│     │   ├── 00000000_cam.txt   
│     │   ├── 00000001_cam.txt   
│     │   └── ...                
│     ├── depth_gt (ground truth depth maps)
│     │   ├── 00000000.pfm   
│     │   ├── 00000001.pfm   
│     │   └── ...                
│     ├── images (images at 7 light indexes) 
│     │   ├── 0 (light index 0)
│     │   │   ├── 00000000.jpg       
│     │   │   ├── 00000001.jpg
│     │   │   └── ...
│     │   ├── 1 (light index 1)
│     │   └── ...                
│     ├── masks (depth map masks) 
│     │   ├── 00000000.png       
│     │   ├── 00000001.png       
│     │   └── ...                
│     └── pair.txt
└── ...
  • In, set MVS_TRAINING as the root directory of the converted dataset; set --output_path as the directory to store the checkpoints.
  • Train the model by running sh
  • The output consists of one checkpoint (model parameters) and one TorchScript module per epoch named as params_<epoch_id>.ckpt and module_<epoch_id>.pt respectively.

Legacy Training

To train directly on the original DTU dataset the legacy training script (using the legacy MVSDataset from datasets/ needs to be called from the script.

  • In, set MVS_TRAINING as the root directory of the original dataset; set --logdir as the directory to store the checkpoints.
  • Uncomment the appropriate section for legacy training and comment out the other entry.
  • Train the model by running sh


--patchmatch_iteration represents the number of iterations of Patchmatch on multi-stages (e.g., the default number 1,2,2 means 1 iteration on stage 1, 2 iterations on stage 2 and 2 iterations on stage 3). --propagate_neighbors represents the number of neighbors for adaptive propagation (e.g., the default number 0,8,16 means no propagation for Patchmatch on stage 1, using 8 neighbors for propagation on stage 2 and using 16 neighbors for propagation on stage 3). As explained in our paper, we do not include adaptive propagation for the last iteration of Patchmatch on stage 1 due to the requirement of photometric consistency filtering. So in our default case (also for our pretrained model), we set the number of propagation neighbors on stage 1 as 0 since the number of iteration on stage 1 is 1. If you want to train the model with more iterations on stage 1, change the corresponding number in --propagate_neighbors to include adaptive propagation for Patchmatch expect for the last iteration.


This project is done in collaboration with "Microsoft Mixed Reality & AI Zurich Lab".

Thanks to Yao Yao for open-sourcing his excellent work MVSNet. Thanks to Xiaoyang Guo for open-sourcing his PyTorch implementation of MVSNet MVSNet-pytorch.