Translucency perception emerges in deep generative representations for natural image synthesis

ABSTRACT

Material perception is essential in planning interactions with the environment. The visual system relies on diagnostic image features to achieve material perception efficiently. However, discovering the features, especially for translucent materials, has been challenging due to the high variability of material appearances under interactions of shape, lighting, and intrinsic materials. Here, we learn a latent space informative of human translucency perception by developing a deep generative network trained to synthesize images of perceptually persuasive material appearances. Without supervision, human-interpretable scene attributes, including object’s shape, material, and body color, spontaneously emerge in the latent space in a scale-specific manner. Critically, the middle-layers of the latent space selectively encode the translucency features correlating with perception, suggesting that translucent impressions are established in the mid-to-low spatial scale features. Our findings illustrate the promising capability of unsupervised learning in finding representative dimensions for materials and discovering perceptually relevant features for visual inference.