Review of: "MLI-NeRF: Multi-Light Intrinsic-Aware Neural Radiance Fields"
Ruggero Pintus · 2025
General assessmentThe paper introduces a novel method for extracting intrinsic image components from scenes captured under various camera poses and lighting conditions.The authors propose MLI-NeRF (Multiple Light Information Intrinsic-aware Neural Radiance Fields), an extension of NeRF that incorporates light position and leverages multi-light and multi-view information to enhance intrinsic decomposition performance.The key innovation is using NeRF to generate pseudo-intrinsic re ectance and shading, which guide the training of intrinsic image decomposition without requiring ground truth data.The method has been evaluated on both synthetic and real-world datasets.The rst general observation is that the proposed method employs a NeRF-like network to extract the scene's geometry, represented in this case as a Signed Distance Function (SDF).This geometry is then used to compute the normal and shadow maps for each camera pose, which, combined with known lighting conditions, are utilized to estimate the Lambertian albedo of the surface.These components (SDF, re ectance, and shading) are integrated with the original images to train a network that re nes the standard NeRF by incorporating constraints derived from the re ectance and shading intrinsic images.This re nement enhances the performance of novel view synthesis (NVS) and relighting under novel light positions while also enabling shading edits within novel views.Computing normal maps, shading maps, and albedo from multi-view acquisitions with known light positions is a well-studied and e ectively solved problem using various approaches outside of NeRFlike methods.Therefore, the strength of this approach lies in Stage 2, where intrinsic images are leveraged to produce a NeRF that excels in novel view synthesis (NVS) and relighting.In fact, Stage 1 could be replaced with any geometry reconstruction method capable of providing a dense surface Qeios qeios.com