PANDORA: Polarization-Aided Neural Decomposition Of Radiance

Akshat Dave, Yongyi Zhao, Ashok Veeraraghavan

Computational Imaging Lab, Rice University

accepted for ECCV 2022

Create a new Anaconda environment using the supplied environment.yml

conda env create -f environment.yml

Unzip this zip file (4.5 GB) into the data folder of the repo directory. The zip file contains real and rendered multi-view polarimetric datasets shown in the paper.

Refer to dataio/Ours.py and data/Mitsuba2.py for pre-processing of real and rendered data respectively.

Run the following command to train geometry and radiance neural representations from multi-view polarimetric images.

python -m train --config configs/real_ceramic_owl.yaml

Config files input through --config describe the parameters required for training. As an example the parameters for real ceramic owl dataset are described in real_ceramic_owl.yaml

Tensorboard logs, checkpoints, arguments and images are saved in the corresponding experiment folder in logs/.

Using the saved arguments from config.yaml and the saved checkpoint such as latest.pt in logs/, novel views can be rendered using the following command.

python -m tools.render_view 
    --config  logs/our_ceramic_owl_v2/config.yaml
    --load_pt logs/our_ceramic_owl_v2/ckpts/latest.pt

Refer to this documentation in neurecon repo for possible camera trajectories. By default first three views used for training are rendered.

Outputs are saved in the corresponding experiment folder in out/. By default,the outputs include surface normal, diffuse radiance, specular radiance and combined radiance for each view along with the estimated roughness.

This repository adapts code or draws inspiration from

• https://github.com/ventusff/neurecon
• https://github.com/yenchenlin/nerf-pytorch
• https://github.com/Fyusion/LLFF
• https://github.com/elerac/polanalyser
• https://github.com/sxyu/svox2

@article{dave2022pandora,
  title={PANDORA: Polarization-Aided Neural Decomposition Of Radiance},
  author={Dave, Akshat and Zhao, Yongyi and Veeraraghavan, Ashok},
  journal={arXiv preprint arXiv:2203.13458},
  year={2022}
}