Fig. 8: Generalization performance on real-world material photographs.
From: Human gloss perception reproduced by tiny neural networks
a, A scatter plot comparing model responses for each image labelled as either matte or glossy based on behavioural data from a previous study11. The x axis represents the single-kernel model, and the y axis represents the three-layer model. Green and pink circles show images consistently rated as matte and glossy, respectively, by all eight observers. The dotted lines indicate the optimal category boundaries that achieve the highest classification accuracies: 91.9% (d′ = 2.79) for the single-kernel model and 71.9% (d′ = 1.08) for the three-layer model. b, Example images correctly classified as glossy by the single-kernel model. c, Example images correctly classified as matte.
