This is the offical code repository for our paper
Plateau-reduced Differentiable Path Tracing [CVPR 2023]
by Michael Fischer and Tobias Ritschel. For more information, make sure to check out the paper and project page.
Clone the repository, then create a new conda environment, activate it and install the dependencies.
git clone https://github.com/mfischer-ucl/prdpt
cd prdpt
conda create -n prdpt python=3.9
conda activate prdpt
pip install mitsuba
pip install matplotlib
pip install torch torchvision torchaudio
You should then be able to run the example in examples/sphere_example.py.
Tested with Python 3.9, PyTorch 2.0 and CUDA 11.7 on Ubuntu 20.04.4 x64 and an NVIDIA RTX3000 series GPU.
This repository uses Mitsuba3 as rendering backbone. The core functionality,
i.e., the importance sampler and kernel functions, reside in utils_fns.py. We prodive the sphere example on top under
examples/sphere_example.py. The main function running the optimization routine is run_optimization.py.
To run your own optimizations, simply provide a Mitsuba scene and extract the information needed
for your use case (as in setup_shadowscene.py). You will also need to adapt the update_fn to update the scene
parameters accordingly, e.g., to apply rotation instead of translation).
If you want to use a different rendering engine, replace the function get_mts_rendering with your own
rendering function. An example of this is provided in the colab notebook under examples/box_position.ipynb, where we use a simple rectangle-renderer.
You can make use of the ctx_args dict to pass data between function calls.
This code is licensed under the MIT license.
If you find our work useful or plan to (re-) use parts of it in your own projects, please include the following citation:
@inproceedings{fischer2023plateau,
title={Plateau-Reduced Differentiable Path Tracing},
author={Fischer, Michael and Ritschel, Tobias},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4285--4294},
year={2023}
}