Reproduce “PBR-NeRF: Inverse Rendering with Physics-Based Neural Fields” 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
Sean Wu, Shamik Basu, Tim Broedermann, Luc Van Gool, Christos Sakaridis
PBR-NeRF introduces a neural inverse rendering framework that jointly estimates scene geometry, materials, and illumination by integrating two novel physics-based priors for the Disney BRDF model. This approach achieves state-of-the-art material estimation, including over 3 PSNR gain in albedo on the NeILF++ dataset, and improves novel view synthesis on DTU by 0.78 mean PSNR while maintaining physical consistency.
Key insights
- Integrating physics-based priors directly into the loss function can effectively constrain the ill-posed inverse rendering problem, leading to more physically plausible and interpretable scene decompositions.
- Explicitly penalizing diffuse contributions in specular-dominant regions (via NDF-weighting) is crucial for accurately disentangling diffuse and specular BRDF components, preventing common "baked-in" artifacts.
- Enforcing energy conservation on the full BRDF helps prevent overestimation of reflected light, which in turn leads to more stable and accurate material and lighting estimations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.