IterMVS (CVPR 2022)

official source code of paper 'IterMVS: Iterative Probability Estimation for Efficient Multi-View Stereo'

Introduction

IterMVS is a novel learning-based MVS method combining highest efficiency and competitive reconstruction quality. We propose a novel GRU-based estimator that encodes pixel-wise probability distributions of depth in its hidden state. Ingesting multi-scale matching information, our model refines these distributions over multiple iterations and infers depth and confidence. Extensive experiments on DTU, Tanks & Temples and ETH3D show highest efficiency in both memory and run-time, and a better generalization ability than many state-of-the-art learning-based methods.

If you find this project useful for your research, please cite:

@misc{wang2021itermvs,
      title={IterMVS: Iterative Probability Estimation for Efficient Multi-View Stereo}, 
      author={Fangjinhua Wang and Silvano Galliani and Christoph Vogel and Marc Pollefeys},
      journal={CVPR},
      year={2022}
}

Installation

Requirements

• python 3.6
• CUDA 10.1

pip install -r requirements.txt

Reproducing Results

• Download pre-processed datasets (provided by PatchmatchNet): DTU's evaluation set, Tanks & Temples and ETH3D benchmark. Each dataset is organized as follows:

root_directory
├──scan1 (scene_name1)
├──scan2 (scene_name2) 
      ├── images                 
      │   ├── 00000000.jpg       
      │   ├── 00000001.jpg       
      │   └── ...                
      ├── cams_1                   
      │   ├── 00000000_cam.txt   
      │   ├── 00000001_cam.txt   
      │   └── ...                
      └── pair.txt  

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

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_MAX 

pair.txt stores the view selection result. For each reference image, 10 best source views are stored in the 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 
...

Evaluation on DTU:

• For DTU's evaluation set, first download our processed camera parameters from here. Unzip it and replace all the old camera files in the folders cams_1 with new files for all the scans.
• In eval_dtu.sh, set DTU_TESTING as the root directory of corresponding dataset, set --outdir as the directory to store the reconstructed point clouds.
• CKPT_FILE is the path of checkpoint file (default as our pretrained model which is trained on DTU, the path is checkpoints/dtu/model_000015.ckpt).
• Test on GPU by running bash eval_dtu.sh. The code includes depth map estimation and depth fusion. The outputs are the point clouds in ply format.
• For quantitative evaluation, download SampleSet and Points from DTU's website. Unzip them and place Points folder in SampleSet/MVS Data/. The structure looks like:

SampleSet
├──MVS Data
      └──Points

In evaluations/dtu/BaseEvalMain_web.m, set dataPath as the 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)
0.373	0.354	0.363

Evaluation on Tansk & Temples:

• In eval_tanks.sh, set TANK_TESTING as the root directory of the dataset and --outdir as the directory to store the reconstructed point clouds.
• CKPT_FILE is the path of checkpoint file (default as our pretrained model which is trained on DTU, the path is checkpoints/dtu/model_000015.ckpt). We also provide our pretrained model trained on BlendedMVS (checkpoints/blendedmvs/model_000015.ckpt)
• Test on GPU by running bash eval_tanks.sh. The code includes depth map estimation and depth fusion. The outputs are the point clouds in ply format.
• For our detailed quantitative results on Tanks & Temples, please check the leaderboards (Tanks & Temples: trained on DTU, Tanks & Temples: trained on BlendedMVS).

Evaluation on ETH3D:

• In eval_eth.sh, set ETH3D_TESTING as the root directory of the dataset and --outdir as the directory to store the reconstructed point clouds.
• CKPT_FILE is the path of checkpoint file (default as our pretrained model which is trained on DTU, the path is checkpoints/dtu/model_000015.ckpt). We also provide our pretrained model trained on BlendedMVS (checkpoints/blendedmvs/model_000015.ckpt)
• Test on GPU by running bash eval_eth.sh. The code includes depth map estimation and depth fusion. The outputs are the point clouds in ply format.
• For our detailed quantitative results on ETH3D, please check the leaderboards (ETH3D: trained on DTU, ETH3D: trained on BlendedMVS).

Evaluation on custom dataset:

• We support preparing the custom dataset from COLMAP's results. The script colmap_input.py (modified based on the script from MVSNet) converts COLMAP's sparse reconstruction results into the same format as the datasets that we provide.
• Test on GPU by running bash eval_custom.sh.

Training

DTU

• Download pre-processed DTU's training set (provided by PatchmatchNet). The dataset is already organized as follows:

root_directory
├──Cameras_1
├──Rectified
└──Depths_raw

• Download our processed camera parameters from here. Unzip all the camera folders into root_directory/Cameras_1.
• In train_dtu.sh, set MVS_TRAINING as the root directory of dataset; set --logdir as the directory to store the checkpoints.
• Train the model by running bash train_dtu.sh.

BlendedMVS

• Download the dataset.
• In train_blend.sh, set MVS_TRAINING as the root directory of dataset; set --logdir as the directory to store the checkpoints.
• Train the model by running bash train_blend.sh.

Acknowledgements

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