Reproduce “Deep Polarization Cues for Single-shot Shape and Subsurface Scattering Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenhao Li, Trung Thanh Ngo, Hajime Nagahara
This research from Osaka University and Hanoi University of Science and Technology introduces a deep learning framework that utilizes single-shot polarization cues to simultaneously estimate the shape and subsurface scattering (SSS) parameters of translucent objects. The method employs novel polarization representations and a reconstruction network, achieving improved accuracy in shape and SSS parameter estimation compared to existing baselines on a new large-scale synthetic dataset.
Key insights
- Polarization cues provide substantial information for disentangling surface reflection from subsurface scattering, leading to more accurate joint estimation of shape and SSS parameters.
- The proposed "max & min polarization representations" are effective explicit cues for reducing diffuse/specular polarization ambiguity in neural networks.
- A reconstruction network using pure BSDF images during training offers a computationally feasible method for high-dimensional image-space supervision of SSS parameters, bypassing the need for complex differentiable polarized renderers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.