Ideal observer analysis and its neural network model:
An application to recovery of structure from motion.

ABSTRACT
Statistical efficiencies were calculated for the recovery of 3D structure from motion in Tanaka and Ishiguchi (2000). In that study, they used stimulus patterns that vertices of a pair of triangles rotating in the 3D space were projected to a 2D plane and they required observers to discriminate the original areas of the triangles in the 3D space. In the present study, it was examined that neural network models could simulate the discriminabilities of the ideal observer on the recovery of 3D structure from motion. Two network models with different architectures learned noise free patterns by using the backpropagation method. Architectures of the two models were as follows: a Single Step Model had three layers and discriminated the areas directly with 2D coordinates of vertices, and another Double Step Model consisted of two sub-networks which recovered 3D structure in Recovery Phase and then discriminated the areas in Discrimination Phase respectively. Then, these networks were tested with noise added patterns. The results of similarity of the Double Step Model to the ideal observer in the discriminabilities suggested that the Double Step Model learned the way to perform the task of the ideal observer.