MVSNet: Depth Inference for Unstructured Multi-view Stereo (ECCV2018, Oral Presentaion)
Clone or download
Latest commit bc6d0ed Dec 13, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
cnn_wrapper Update README Jul 5, 2018
doc add figures for point cloud output Aug 31, 2018
mvsnet buffer the input and speed up the training/testing (+30% faster in my… Dec 13, 2018
tools Update README Jul 5, 2018
LECENSE MIT license Jul 26, 2018 Fix training camera Sep 19, 2018
requirements.txt tensorflow version Jul 13, 2018



MVSNet is a deep learning architecture for depth map inference from unstructured multi-view images. If you find this project useful for your research, please cite:

  title={MVSNet: Depth Inference for Unstructured Multi-view Stereo},
  author={Yao, Yao and Luo, Zixin and Li, Shiwei and Fang, Tian and Quan, Long},
  journal={European Conference on Computer Vision (ECCV)},

How to Use


  • Check out the source code git clone
  • Install cuda 9.0, cudnn 7.0 and python 2.7
  • Install Tensorflow and other dependencies by sudo pip install -r requirements.txt


  • Download the preprocessed DTU training data (Fixed training cameras, Sep. 19), and upzip it as the MVS_TRANING folder.
  • Enter the MVSNet/mvsnet folder, in, set dtu_data_root to your MVS_TRANING path.
  • 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 correspondingly.
  • Train the network python


  • 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 MVSNet model and upzip it as MODEL_FOLDER.
  • Enter the MVSNet/mvsnet folder, in, set pretrained_model_ckpt_path to MODEL_FOLDER/model.ckpt
  • Test on this dataset python --dense_folder TEST_DATA_FOLDER.
  • Inspect the .pfm format outputs in TEST_DATA_FOLDER/depths_mvsnet using python .pfm. For example the depth map and probability map for image 00000012 should look like:
reference image depth map probability map


MVSNet itself only produces per-view depth maps. To generate the 3D point cloud, we need to apply depth map filter/fusion for post-processing. As our implementation of this part is depended on the Altizure internal library, currently we could not provide the corresponding code. Fortunately, depth map filter/fusion is a general step in MVS reconstruction, and there are similar implementations in other open-source MVS algorithms. We provide the script to utilize fusibile for post-processing (thank Silvano Galliani for the excellent code!).

To run the post-processing:

  • Check out the modified version fusibile git clone
  • Install fusibile by cmake . and make, which will generate the executable at FUSIBILE_EXE_PATH
  • Run post-processing python --dense_folder TEST_DATA_FOLDER --fusibile_exe_path FUSIBILE_EXE_PATH
  • The final point cloud is stored in TEST_DATA_FOLDER/points_mvsnet/consistencyCheck-TIME/final3d_model.ply.

We observe that the point cloud output of is very similar to our own implementation. For detailed differences, please refer to MVSNet paper and Galliani's paper. The point cloud for scan9 should look like:

point cloud result ground truth point cloud

File Formats

Each project folder should contain the following

├── images                 
│   ├── 00000000.jpg       
│   ├── 00000001.jpg       
│   └── ...                
├── cams                   
│   ├── 00000000_cam.txt   
│   ├── 00000001_cam.txt   
│   └── ...                
└── pair.txt               

If you want to apply MVSNet to your own data, please structure your data into such a folder.

Image Files

All image files are stored in the images folder. We index each image using an 8 digit number starting from 00000000. The following camera and output files use the same indexes as well.

Camera Files

The camera parameter of one image is stored in a cam.txt file. The text file contains the camera extrinsic E = [R|t], intrinsic K and the depth range:

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


Note that the depth range and depth resolution are determined by the minimum depth DEPTH_MIN, the interval between two depth samples DEPTH_INTERVAL, and also the depth sample number (max_d in the training/testing scripts). The max_d is left in the scripts for users to flexibly control the depth range. We also left the interval_scale for controlling the depth resolution. The maximum depth is then computed as:

DEPTH_MAX = DEPTH_MIN + (interval_scale * DEPTH_INTERVAL) * (max_d - 1)

View Selection File

We store the view selection result in the pair.txt. For each reference image, we calculate its view selection scores with each of the other views, and store the 10 best views in the pair.txt 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 

Output Format

The script will create a depths_mvsnet folder to store the running results, including the depth maps, probability maps, scaled/cropped images and the corresponding cameras. The depth and probability maps are stored in .pfm format. We provide the python IO for pfm files in the script, and for the c++ IO, we refer users to the Cimg library. To inspect the pfm format results, you can simply type python .pfm.