Monocular Depth Decomposition of Semi-Transparent Volume Renderings

This is the official PyTorch code for our Paper.

Project Page | Paper | Pre-Print (Open Access)

If you find our work useful, please cite our paper:

@article{engel2023stdepth,
  title={Monocular Depth Decomposition of Semi-Transparent Volume Renderings},
  author={Engel, Dominik and Hartwig, Sebastian and Ropinski, Timo},
  journal={IEEE Transactions on Visualization and Computer Graphics},
  year={2023},
  doi={10.1109/TVCG.2023.3245305},
}

`mono-depth-estimation` Repository

The depth estimation networks are implemented in a submodule, so you should git submodule update --init after cloning this repo. Big thanks to Sebastian Hartwig (@kopetri) who did most of the work in compiling those methods into one repository.

Pre-Trained weights

TODO

Setup Docker Container

For Training:

docker run -it --gpus all --ipc=host -v $PWD:$PWD -w $PWD pytorch/pytorch:1.9.1-cuda11.1-cudnn8-runtime bash
apt-get install -y libsm6 libxext6 libxrender-dev libglib2.0-0 libgl1-mesa-glx git python3-opencv
pip install -r mono-depth-estimation/requirements.txt
python mono-depth-estimation/train.py ....

For inference and using the notebooks from this repo:

docker run -it --gpus all --ipc=host -v $PWD:$PWD -w $PWD -p 9999:9999 pytorch/pytorch:1.9.1-cuda11.1-cudnn8-runtime bash
apt-get install -y libsm6 libxext6 libxrender-dev libglib2.0-0 libgl1-mesa-glx git python3-opencv
pip install -r requirements.txt
pip install -r mono-depth-estimation/requirements.txt
jupyter notebook --ip 0.0.0.0 --port 9999 --allow-root

If your data lives outside of this project folder, additionally mount your data folders to the containers (-v /path/to/data:/path/inside/container).

Run Training

To run the training, use the mono-depth-estimation/train.py script. Using --help you should get an overview of possible arguments. We recommend using the above Docker setup to run those commands. Paths will have to be adjusted accordingly. For just training depth prediciton start with this sample command (mono-depth-estimation on branch stdepthmono):

python mono-depth-estimation/train.py --min_epochs 10 --max_epochs 50 --gpus -1 --worker 4 bts --learning_rate 1e-5 --batch_size 16 --loss bts stdepth --background-depth-max --depth-method wysiwyp --path '/mnt/hdd/dome/STDepth_data4' --training --validation

For layered prediction start with this sample (mono-depth-estimation on branch stdepthmulti):

python mono-depth-estimation/train.py --name BTS_ours --precision 32 --gpus -1 bts --image-residuals --loss silma+composite+colorssim+fbdivergence --learning_rate 2e-5 --batch_size 32 stdepthmulti2 --single-layer --path '/mnt/hdd/dome/STDepth_data_multi_wys2' --training --validation

The parameters like loss weights default to the values reported in the paper.

Relevant notebooks

change_viewpoint.ipynb demonstrates novel view synthesis, creation of stereo pairs and animations
evaluate_depthmaps.ipynb demonstrates network inference and evaluation by comparing with ground truth
generate_dataset.ipynb shows generation of the dataset for training WITHOUT layered representations
generate_multidepth.ipynb shows generation of the dataset for training of layered representations
generate_multidepth_bg.ipynb further includes creating black/white-ish backgrounds with shot noise
predict_image.ipynb and predict_vessels.ipynb demonstrates network inference on images just given as jpeg or png without any ground truth. This includes many of the presented applications, like ambient occlusion, contour and chromadepth shading.
visualize_depth_techniques.ipynb creates the visualization of the different depth techniques shown in our paper

viscom-ulm/MonocularDepthDecomposition