By Hong-Xing Yu*, Michelle Guo*, Alireza Fathi, Yen-Yu Chang, Eric Ryan Chan, Ruohan Gao, Thomas Funkhouser, Jiajun Wu

arXiv link: https://arxiv.org/abs/2303.06138

git clone https://github.com/michguo/osf.git
cd osf
conda env create -f environment.yml

You can download the data and pretrained models from the following Google Drive links:

• Data (2.0 GB): https://drive.google.com/file/d/1IWNt5R2Mpp1XHo8JiwfHUZXI7Coi4OXO/view?usp=share_link
• Models (0.8 GB): https://drive.google.com/file/d/1oj-F4SWgt-vfRyS9MHIVf9cdJnZe-p81/view?usp=share_link

Configuration files can be found in the configs folder. To train an OSF, run the following command:

python run_osf.py --config ${CONFIG_PATH}

For testing, you can run

python run_osf.py --config ${CONFIG_PATH} --render_only --render_test

After training individual OSFs, you can compose them together into arbitrary scene arrangements at test time. This example composes the checkers background and the ObjectFolder objects together into a scene:

python run_osf.py --config=configs/compose.txt

For the extension of KiloOSF where we distill a trained OSF model to accelerate rendering, please refer to KiloOSF.

@article{yu2023osf,
    title={Learning Object-centric Neural Scattering Functions for Free-viewpoint Relighting and Scene Composition},
    author={Yu, Hong-Xing and Guo, Michelle and Fathi, Alireza and Chang, Yen-Yu and Chan, Eric Ryan and Gao, Ruohan and Funkhouser, Thomas and Wu, Jiajun},
    journal={Transactions on Machine Learning Research},
    year={2023}
}

Our code framework is adapted from nerf-pytorch.