The abstract is read at the publisher; the summary is JClub's.
Deep neural networks trained to estimate reflectance from rendered 3D scenes show similar reflectance performance and susceptibility to classic lightness illusions across ResUNet, VGGUNet, MobileUNet, and Dense Prediction Transformer.
- Why it matters: Lightness constancy requires representations correlated with surface reflectance despite illumination and context changes, and it is unclear whether deep models’ lightness behavior depends on specific architectural or training choices.
- What they did: The study trained multiple architectures to estimate reflectance from rendered images of 3D scenes, varying whether initial stimulus-encoding used pre-trained weights or whether all weights were optimized for reflectance estimation, then tested held-out performance and saliency.
- The result: All tested architectures and training procedures produced similar held-out reflectance estimation and showed classic lightness illusions, with saliency analyses indicating broadly similar spatial sensitivity, implying these properties robustly emerge from reflectance inference optimization.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.