A new paper identifies a significant flaw in how intrinsic image decomposition models are evaluated, specifically on the MPI Sintel dataset. Researchers found that splitting datasets by frames, rather than by scenes, leads to inflated performance metrics by up to 2.0 dB due to spatial similarity between training and testing frames. They propose using scene-level splits as the standard and demonstrate a new physics-informed decomposition method with source-separable uncertainty, which improves downstream tasks by filtering uncertain pixels. AI
Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →
IMPACT Highlights the need for more rigorous evaluation protocols in computer vision, potentially impacting how future decomposition models are benchmarked and developed.
RANK_REASON The cluster contains an academic paper detailing a new evaluation protocol and a novel method for intrinsic image decomposition.