Fig. 7: Results of a generalization test using Serrano dataset.
From: Human gloss perception reproduced by tiny neural networks
a,b, Nine lighting environments (a) and shapes (b) used to generate rendered images. c, Pearson’s correlation coefficient between the model response of the single-kernel model trained by all 3,888 images and perceptual rating measured in crowd-source experiment47 for each lighting and shape. Each bar shows the correlation across 4,680 images. d, The case for a three-layer model trained by all 3,888 images, which shows slightly lower generalization ability compared with the single-kernel model.
