Home

Awesome

MVSNet & R-MVSNet

<font color="red"> [News] BlendedMVS dataset is released!!!</font> (project link).

About

MVSNet is a deep learning architecture for depth map inference from unstructured multi-view images, and R-MVSNet is its extension for scalable learning-based MVS reconstruction. If you find this project useful for your research, please cite:

@article{yao2018mvsnet,
  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)},
  year={2018}
}
@article{yao2019recurrent,
  title={Recurrent MVSNet for High-resolution Multi-view Stereo Depth Inference},
  author={Yao, Yao and Luo, Zixin and Li, Shiwei and Shen, Tianwei and Fang, Tian and Quan, Long},
  journal={Computer Vision and Pattern Recognition (CVPR)},
  year={2019}
}

If BlendedMVS dataset is used in your research, please also cite:

@article{yao2020blendedmvs,
  title={BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks},
  author={Yao, Yao and Luo, Zixin and Li, Shiwei and Zhang, Jingyang and Ren, Yufan and Zhou, Lei and Fang, Tian and Quan, Long},
  journal={Computer Vision and Pattern Recognition (CVPR)},
  year={2020}
}

How to Use

Installation

Download

Training

Validation

Testing

<img src="doc/image.png" width="250"><img src="doc/depth_example.png" width="250"><img src="doc/probability_example.png" width="250">
reference imagedepth mapprobability map

Post-Processing

R/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 depthfusion.py to utilize fusibile for post-processing (thank Silvano Galliani for the excellent code!).

To run the post-processing:

We observe that depthfusion.py produce similar but quantitatively worse result to our own implementation. For detailed differences, please refer to MVSNet paper and Galliani's paper. The point cloud for scan9 should look like:

<img src="doc/fused_point_cloud.png" width="375"><img src="doc/gt_point_cloud.png" width="375">
point cloud resultground truth point cloud

Reproduce Paper Results

The following steps are required to reproduce depth map/point cloud results:

R-MVSNet point cloud results with full post-processing are also provided: DTU evaluation point clouds

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 R/MVSNet to your own data, please structure your data into such a folder. We also provide a simple script colmap2mvsnet.py to convert COLMAP SfM result to R/MVSNet input.

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:

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

intrinsic
K00 K01 K02
K10 K11 K12
K20 K21 K22

DEPTH_MIN DEPTH_INTERVAL (DEPTH_NUM DEPTH_MAX) 

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 DEPTH_NUM (or max_d in the training/testing scripts if DEPTH_NUM is not provided). 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:

TOTAL_IMAGE_NUM
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 
...

MVSNet input from SfM output

We provide a script to convert COLMAP SfM result to R/MVSNet input. After recovering SfM result and undistorting all images, COLMAP should generate a dense folder COLMAP/dense/ containing an undistorted image folder COLMAP/dense/images/ and an undistorted camera folder COLMAP/dense/sparse/. Then, you can apply the following script to generate the R/MVSNet input:

python colmap2mvsnet.py --dense_folder COLMAP/dense

The depth sample number will be automatically computed using the inverse depth setting. If you want to generate the MVSNet input with a fixed depth sample number (e.g., 256), you could specified the depth number via --max_d 256.

Output Format

The test.py 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 preprocess.py script, and for the c++ IO, we refer users to the Cimg library. To inspect the pfm format results, you can simply type python visualize.py .pfm.

Changelog

2020 April 13

2020 March 2

2020 Feb 29

2019 April 29

2019 April 10

2019 March 14

2019 March 11

2019 March 7

2019 March 1

2019 Feb 28