SF-MVSNet

🔧# Installation

conda create -n gomvs python=3.8 # use python 3.8
conda activate gomvs
pip install -r requirements.txt
#Our code is trained and tested on NVIDIA RTX 3090 GPU (with python=3.8, torch=1.12.1 cuda=11.3)

If you are using NVIDIA RTX 3090 GPU too, please use the following commands reinstall pytorch

# uninstall
pip uninstall pytorch

# install
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
# or
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

Dataset

We use the dataset GoMVS provided, which is the same for many other MVSNet methods except normals. Please refer to GoMVS for download link.

Run

Note SF-MVSNet require roughly 20GB+ GPU memory for 1 batchsize, and We didn't test finetune.py because our goal is to test TNT dataset without fine-tuning.

Train

Please specify MVS_TRAINING and NORMAL_PATH in scripts/train/*.sh, then run the following bash to train on DTU dataset

bash scripts/train/train_dtu.sh

Evaluation

Pretrained model

Pretrained model can be downloaded in this link, verify code is t52u.

DTU Dataset

For quantitative results:

1. build fusibile from this repo. Instructions are in MVSNet Post-Processing section.
2. run bash scripts/test/test_dtu.sh to get depth map and point clouds.
3. run DTU matlab code.

TNT Dataset

Run the following bash to get depth map and point clouds.

# get depth map
bash scripts/test/test_tnt.sh

# get point clouds
bash scripts/test/dynamic_fusion.sh

Acknowledgments

We borrow the code from TransMVSNet,GoMVS, We express gratitude for their marvelous contributions!