The power and versatility of the human visual system derives in large part from its amazing ability to find structure and organization in the images encoded by the retinas. To discover and describe structure, the visual system uses a wide array of perceptual grouping/segregation mechanisms. This sophisticated array of mechanisms is absolutely essential for human ability to recognize objects and correctly interpret visual scenes. During the previous funding period, we have taken a systematic quantitative approach to the study of the perceptual grouping mechanisms. We propose to continue and extend our general approach, which consists of three major parts: (1) measuring the statistical properties of the visual images that are relevant for perceptual grouping, (2) developing models of perceptual grouping that are informed by the statistical properties of visual images, by the physiology and psychophysics of low-level vision, and by computational principles, and (3) testing the predictions of these models and competing models in psychophysical experiments. We propose two methods for measuring image statistics. One method is to extract local features from images and then compute simple co- occurrence statistics; e.g., the joint probabilities of all possible geometrical relationships between pairs of local edge elements extracted from representative collections of natural images. The other method is to extract local features from images and then use an image- tracing procedure to measure the Bayesian co-occurrence statistics; e.g., the likelihood that any given pair of edge elements belong to the same physical contour versus different physical contours. Most of the proposed modeling and psychophysical work will focus on the mechanisms of contour grouping and motion grouping. The contour grouping experiments are directed at testing and extending our successful model of contour grouping based upon natural image statistics. The motion grouping experiments will examine the role of motion information in contour grouping and role of spatial information in motion grouping. We also plan to test motion-grouping models that we will develop from measurements of natural video image statistics.
Showing the most recent 10 out of 34 publications
