bioRxiv · 20 Nov 2026
Lightness illusions and saliency maps for reflectance estimation are similar across a wide range of deep neural network architectures
Jaykishan Patel, Alban Flachot, Javier Vazquez-Corral, David H. Brainard, Thomas S. A. Wallis, Richard F Murray
Citations by database
Each database indexes a different set of journals, preprints and books, so they are shown side by side rather than blended.
- OpenCitations0
- Crossref0
- Europe PMC0
Citations per year
The current year is still counting.
- 2026 so far0
Abstract
People perform many visual tasks that depend on having representations that remain correlated with surface reflectance despite large variations in illumination and context. This ability is known as `lightness constancy'. Developing image-computable models of lightness constancy has been challenging, but we recently showed that deep neural networks trained for reflectance estimation achieve good performance on some lightness tasks, and furthermore are susceptible to many of the same lightness illusions as human observers. Here we extend this work by testing whether these findings are specific to the network architecture we used in our previous work, or whether they generalize broadly across architectures and training procedures. We evaluated a range of architectures (ResUNet, VGGUNet, MobileUNet, and Dense Prediction Transformer) that we trained to estimate reflectance from rendered images of three-dimensional scenes. We compared networks where the initial, stimulus-encoding phase used pre-trained weights learned for other tasks, and networks where all weights were optimized for reflectance estimation. We found that all architectures and training procedures tested produced similar reflectance estimation performance on held-out images, and were susceptible to several classic lightness illusions. Furthermore, saliency analyses showed that all networks had broadly similar spatial sensitivity across input images. Together, these results suggest that lightness constancy and susceptibility to lightness illusions do not depend strongly on details of network architecture and training, but instead often emerge robustly in networks that have been optimized to infer surface reflectance from luminance.
What the data shows
- No database records a citation yet, as is usual for a paper this new.
Graph
This paper in the middle, its authors and their most-cited papers below, the papers it cites to the left and the papers citing it to the right. The timeline lays every paper out by year.
Select a node for its details. Drag to pan, scroll to zoom, drag a node to move it.
Start with
Cited by 0 papers
Who cites this paper, from open citation data, with Crossref's record of each; the 100 newest when there are more.
References (0)
What it cites, from open citation data, with Crossref's record of each.