Fig. 3: Summary of computational model behaviours.
From: Human gloss perception reproduced by tiny neural networks
Dark-green circles show individual human observers (N = 295), where the x axis and the y axis show the correlation over 74 images to the physical ground truth and the correlation to the average of the rest of the observers, respectively. Other data points show candidate computational models. Grey diamonds show low-level luminance image statistics models, and the grey star shows a multiple regression model based on the luminance statistics (best fit to human data). Light-green downward triangles show models based on statistics computed from specular reflection images (sharpness, coverage and sub-band contrast), and the light-green star symbol shows a multiple regression based on the specular metrics (best fit to human data). Triangles and squares show one-layer models with different kernel numbers (1, 2, 4 and 9; 9, 16, 32 and 64), where blue symbols refer to the networks trained on images labelled by observer judgements and pink symbols show networks trained on images labelled by physical ground-truth labels. The magenta rightward triangle shows ResNet18 trained on physical ground-truth labels. Orange symbols represent CNNs trained with additional images, generated using novel shapes and lighting conditions not included in the 3,888 images. Orange squares correspond to the three-layer model with 64 filters, and orange rightward triangles show ResNet18. The orange numbers on the plot show the number of additional images used.
