conda create -n venv python=3.6
conda activate venv
conda install pytorch==1.1.0 torchvision cudatoolkit=9.0 -c pytorch
pip install tensorboardX
pip install opencv-python
pip install plyfile
Your CUDA version should be 9.0.
cd prior-mvsnet/MYTH
python setup.py install
Our model is trained on DTU dataset. All training parameters are configured in file config.json.
To train model, first we need to prepare prior depth&confidence from CasMVSNet. Then, run this command to train model:
python train.py --config config.json
First, get the prior depth from CasMVSNet.
All outputs are saved in a folder casmvs_outputs which is placed in the test root folder.
Then, run our model to predict depth & confidence
python test.py --dataset general_eval --batch_size 1 --testpath <your test root folder> --testlist <your dataset name> --resume <pretrained model> --outdir outputs --interval_scale 0.8 --num_view 5 --depth_scale 0.001
By default, our pretrained model is tested with number of depth planes [80, 32, 8] and the depth interval ratio [4, 2, 1].
You can change these parameters in file pretrained/full/config.json.
Note that you need to change some parameters to fit to your dataset such as depth_scale, num_view, max_h, max_w or total of depth planes numdepth.
The depth scale depth_scale should be in milimeter (minimum depth is around 400mm).
Examples of our evaluation datasets
# with the Family scene of Tanks&Temples dataset,
python test.py --dataset general_eval --batch_size 1 --testpath /mnt/sdb/khang/tanksandtemples/intermediate --testlist Family --resume pretrained/full/model_best.pth --outdir outputs --interval_scale 0.8 --num_view 7 --depth_scale 0.0006 --numdepth 320 --max_h 1080 --max_w 1920
# with DTU dataset, change #depth planes to [64, 32, 8]
python test.py --dataset general_eval --batch_size 1 --testpath /mnt/sdb/khang/dtu_dataset/test --testlist lists/dtu/test.txt --resume pretrained/full/model_best.pth --outdir outputs --interval_scale 0.8 --num_view 7 --depth_scale 1.0 --numdepth 256 --max_h 864 --max_w 1152