Visual computational model for viewpoint dependency in human object recognition

摘要

In a chain of psychophysical and physiological studies about cognition of three-dimensional objects by only two-dimensional information from a specific viewpoint, it has not been clear how the two- and three-dimensional percepts, i.e., viewer-centerd and object-centered representations, are processed in our brain. In the present study we take notice of the diversity level of human object recognition performance when changing the viewer position, i.e., viewpoint dependency, and discuss about the mechanism of visual cognition from a computational perspective. The task in our psychophysical experiments was discrimination of unfamiliar objects by their images seen from some viewpoints. We found that the effect of high-frequency components in such stimuli for viewpoint dependency was different between object groups with distinct volumetric features. Based on the results, we proposed a mathematical model in which both mechanisms processing viewer- and object-centered representations are integrated.