Biological visual systems make use of many different sources of information (""""""""cues"""""""") for visual judgments. For depth and shape estimation, for example, these include occlusion, texture, perspective, motion parallax, disparity, shading and contour. The combination of these cues is based on the relative reliabilities of the individual cues, but cannot occur until cues are promoted to a commensurate scale by filling in one or more needed parameters (e.g., the fixation distance and azimuth for depth and slant estimates). These parameters are also estimated using multiple cues (e.g., both retinal and oculomotor cues for the viewing geometry). We propose statistical decision theoretic models for ideal behavior in the visual estimation of scene properties and for movement planning. The ideal observer or actor must take into account measurement uncertainty, associated with different outcomes, and prior information about the current state of the world. We propose experiments intended to clarify how human observers promote and combine cues for vision and for the visual control of action. The experimental methods used are based on perturbation analysis which permits examination of a system that can potentially react to distortions and inconsistencies in the stimuli. The proposed research consists of three major tasks. (1) We will analyze observer behavior relative to predictions of ideal Bayesian decision makers confronted by the same levels of uncertainty in tasks of perceptual decision, reaching and grasping. (2) We will examine cue combination in the service of cue promotion, again with reference to ideal behavior. (3) We will continue our studies of spatial interpolation performance so as to better understand such aspects of the underlying model as the prior distribution, and the methods used by the observer to be statistically robust (which, in this context, is closely related to the scene segmentation problem).
Showing the most recent 10 out of 113 publications
