Fig. 5.
The color statistics of objects predict the average surprisal of colors. Objects in the Microsoft Research Asia database of 20,000 photographs were identified by human observers who were blind to the purpose of our study (31). The colors of the pixels in the images were binned into the 80 colors defined by the Munsell chips used in the behavioral experiments (across the images there were 9.2 × 108 object pixels and 1.54 × 109 background pixels). The y axis shows the [(number of pixels of given color in objects)/(number of pixels of given color in objects + number of pixels of given color in backgrounds)]; the color chip ranking is that obtained for the Tsimane'. Error bars are SE. The three languages were not significantly different from each other (English: slope = −0.0064, ρ = −0.57, P value = 3 × 10−8; Bolivian-Spanish: slope = −0.0049, ρ = −0.44, P value = 5 × 10−4; Tsimane': slope = −0.0054, ρ = −0.49, P value = 5 × 10−6).