Reproduce “Evidential Neural Radiance Fields” on OpenResearch

About this paper

Ruxiao Duan, Alex Wong

A probabilistic framework integrates evidential deep learning into Neural Radiance Fields, enabling the quantification of both aleatoric and epistemic uncertainties from a single model and forward pass. This approach yields high-fidelity scene reconstructions with competitive uncertainty estimation, supporting applications like scene cleaning and active learning while maintaining computational efficiency.

Key insights

• Evidential Deep Learning can be effectively adapted to the hierarchical, volumetric rendering process of NeRFs to simultaneously model both aleatoric and epistemic uncertainties.
• Uncertainty estimation can be achieved from a single network and forward pass without sacrificing rendering quality or computational efficiency, addressing a critical trade-off.
• Distinct aleatoric and epistemic uncertainties provide valuable signals for identifying data noise versus model knowledge gaps, enabling applications like geometric artifact removal and active learning.

Summary by alphaXiv.