A deep convolutional neural network trained for lightness constancy is susceptible to lightness illusions

Human viewers are able to perform tasks that depend on accurate estimates of surface reflectance, even across large changes in illumination and context. This is a remarkable ability, and successful image-computable models of how the visual system achieves this have remained elusive. Recently, deep convolutional neural networks (CNNs) have been developed that are adept at estimating surface reflectance. Here we evaluated one such network as a starting point for a new model of human lightness perception by testing whether it was susceptible to a range of classic lightness illusions. We implemented a CNN and trained it via supervised learning to estimate surface reflectance at each pixel in grayscale, rendered images of geometric objects. We examined the network’s output on several illusions, including the argyle, Koffka, snake, simultaneous contrast, White’s, and checkerboard illusions, as well as control figures. We included variants where low-luminance regions important to the illusions were generated either by low reflectance or by cast shadows. For comparison, we carried out a lightness matching experiment with human observers using the same stimuli, and also examined the outputs of three classic lightness and brightness models. The CNN largely removed lighting effects such as shading and shadows, and produced good reflectance estimates on a test set. It also qualitatively predicted the illusions perceived by humans in most cases, the exceptions being White’s and checkerboard illusions. The CNN outperformed classical models, both at estimating reflectance and at tracking human lightness matches. These findings support a normative view of lightness perception and highlight the promise of deep learning models in this area.