Fig. 7. The necessity of texture information in inferring objects’ size in DCNNs.
a Images from the ImageNet were transformed with various painting styles, where the shape information was preserved and the texture information distorted. b The RSM of Conv4’s response to objects’ size is similar to the ideal observer. c Only the fourth axis of object space constructed by the stylized AlexNet encoded the size feature. d Logarithm correspondence between PC4 and objects’ size showed much less sensitivity to texture as compared to shape.