Tips: If you meet any problems when reproduce our results, please contact Yikang Ding ([email protected]). We are happy to help you solve the problems and share our experience.

• 09.2022: Add more detailed instruction of how to reproduce the reported results (see testing-on-dtu).
• 09.2022: Fix the bugs in MATLAB evaluation code (remove the debug code).
• 09.2022: Fix the bug of default fuse parameters of gipuma, which could have a great impact on the final results.
• 09.2022: Update the website link and instruction of installing gipuma, which would affect the fusion quality.

In this paper, we present TransMVSNet, based on our exploration of feature matching in multi-view stereo (MVS). We analogize MVS back to its nature of a feature matching task and therefore propose a powerful Feature Matching Transformer (FMT) to leverage intra- (self-) and inter- (cross-) attention to aggregate long-range context information within and across images. To facilitate a better adaptation of the FMT, we leverage an Adaptive Receptive Field (ARF) module to ensure a smooth transit in scopes of features and bridge different stages with a feature pathway to pass transformed features and gradients across different scales. In addition, we apply pair-wise feature correlation to measure similarity between features, and adopt ambiguity-reducing focal loss to strengthen the supervision. To the best of our knowledge, TransMVSNet is the first attempt to leverage Transformer into the task of MVS. As a result, our method achieves state-of-the-art performance on DTU dataset, Tanks and Temples benchmark, and BlendedMVS dataset.

Our code is tested with Python==3.6/3.7/3.8, PyTorch==1.6.0/1.7.0/1.9.0, CUDA==10.2 on Ubuntu-18.04 with NVIDIA GeForce RTX 2080Ti. Similar or higher version should work well.

To use TransMVSNet, clone this repo:

git clone https://github.com/MegviiRobot/TransMVSNet.git
cd TransMVSNet

We highly recommend using Anaconda to manage the python environment:

conda create -n transmvsnet python=3.6
conda activate transmvsnet
pip install -r requirements.txt

We also recommend using apex, you can install apex from the official repo.

In TransMVSNet, we mainly use DTU, BlendedMVS and Tanks and Temples to train and evaluate our models. You can prepare the corresponding data by following the instructions below.

For DTU training set, you can download the preprocessed DTU training data and Depths_raw (both from Original MVSNet), and unzip them to construct a dataset folder like:

dtu_training
 ├── Cameras
 ├── Depths
 ├── Depths_raw
 └── Rectified

For DTU testing set, you can download the preprocessed DTU testing data (from Original MVSNet) and unzip it as the test data folder, which should contain one cams folder, one images folder and one pair.txt file.

We use the low-res set of BlendedMVS dataset for both training and testing. You can download the low-res set from orignal BlendedMVS and unzip it to form the dataset folder like below:

BlendedMVS
 ├── 5a0271884e62597cdee0d0eb
 │     ├── blended_images
 │     ├── cams
 │     └── rendered_depth_maps
 ├── 59338e76772c3e6384afbb15
 ├── 59f363a8b45be22330016cad
 ├── ...
 ├── all_list.txt
 ├── training_list.txt
 └── validation_list.txt

Download our preprocessed Tanks and Temples dataset and unzip it to form the dataset folder like below:

tankandtemples
 ├── advanced
 │  ├── Auditorium
 │  ├── Ballroom
 │  ├── ...
 │  └── Temple
 └── intermediate
        ├── Family
        ├── Francis
        ├── ...
        └── Train

Set the configuration in scripts/train.sh:

• Set MVS_TRAINING as the path of DTU training set.
• Set LOG_DIR to save the checkpoints.
• Change NGPUS to suit your device.
• We use torch.distributed.launch by default.

To train your own model, just run:

bash scripts/train.sh

You can conveniently modify more hyper-parameters in scripts/train.sh according to the argparser in train.py, such as summary_freq, save_freq, and so on.

For a fair comparison with other SOTA methods on Tanks and Temples benchmark, we finetune our model on BlendedMVS dataset after training on DTU dataset.

Set the configuration in scripts/train_bld_fintune.sh:

• Set MVS_TRAINING as the path of BlendedMVS dataset.
• Set LOG_DIR to save the checkpoints and training log.
• Set CKPT as path of the loaded .ckpt which is trained on DTU dataset.

To finetune your own model, just run:

bash scripts/train_bld_fintune.sh

For easy testing, you can download our pre-trained models and put them in checkpoints folder, or use your own models and follow the instruction below.

Important Tips: to reproduce our reported results, you need to:

• compile and install the modified gipuma from Yao Yao as introduced below
• use the latest code as we have fixed tiny bugs and updated the fusion parameters
• make sure you install the right version of python and pytorch, use some old versions would throw warnings of the default action of align_corner in several functions, which would affect the final results
• be aware that we only test the code on 2080Ti and Ubuntu 18.04, other devices and systems might get slightly different results
• make sure that you use the model_dtu.ckpt for testing

To start testing, set the configuration in scripts/test_dtu.sh:

• Set TESTPATH as the path of DTU testing set.
• Set TESTLIST as the path of test list (.txt file).
• Set CKPT_FILE as the path of the model weights.
• Set OUTDIR as the path to save results.

Run:

bash scripts/test_dtu.sh

Note: You can use the gipuma fusion method or normal fusion method to fuse the point clouds. In our experiments, we use the gipuma fusion method by default. With using the uploaded ckpt and latest code, these two fusion methods would get the below results:

To install the gipuma, clone the modified version from Yao Yao. Modify the line-10 in CMakeLists.txt to suit your GPUs. Othervise you would meet warnings when compile it, which would lead to failure and get 0 points in fused point cloud. For example, if you use 2080Ti GPU, modify the line-10 to:

set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-O3 --use_fast_math --ptxas-options=-v -std=c++11 --compiler-options -Wall -gencode arch=compute_70,code=sm_70)

If you use other kind of GPUs, please modify the arch code to suit your device (arch=compute_XX,code=sm_XX). Then install it by cmake . and make, which will generate the executable file at FUSIBILE_EXE_PATH. Please note

For quantitative evaluation on DTU dataset, download SampleSet and Points. Unzip them and place Points folder in SampleSet/MVS Data/. The structure looks like:

SampleSet
├──MVS Data
      └──Points

In DTU-MATLAB/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 DTU-MATLAB/BaseEvalMain_web.m in matlab.

We also upload our final point cloud results to here. You can easily download them and evaluate them using the MATLAB scripts, the results look like:

We recommend using the finetuned models (model_bld.ckpt) to test on Tanks and Temples benchmark.

Similarly, set the configuration in scripts/test_tnt.sh:

• Set TESTPATH as the path of intermediate set or advanced set.
• Set TESTLIST as the path of test list (.txt file).
• Set CKPT_FILE as the path of the model weights.
• Set OUTDIR as the path to save resutls.

To generate point cloud results, just run:

bash scripts/test_tnt.sh

Note that：

• The parameters of point cloud fusion have not been studied thoroughly and the performance can be better if cherry-picking more appropriate thresholds for each of the scenes.
• The dynamic fusion code is borrowed from AA-RMVSNet.

For quantitative evaluation, you can upload your point clouds to Tanks and Temples benchmark.

@inproceedings{ding2022transmvsnet,
  title={Transmvsnet: Global context-aware multi-view stereo network with transformers},
  author={Ding, Yikang and Yuan, Wentao and Zhu, Qingtian and Zhang, Haotian and Liu, Xiangyue and Wang, Yuanjiang and Liu, Xiao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={8585--8594},
  year={2022}
}

We borrow some code from CasMVSNet, LoFTR and AA-RMVSNet. We thank the authors for releasing the source code.