Binding the pieces: Efficacies of grouping cues

Abstract

When a newborn baby sees the booming, buzzing confusion of the visual world, how does it begin to piece together the parts that make up contours, objects, and background? The Gestaltists have identified many of the cues to which the adult visual system is sensitive, but it is not known which of these cues are innately specified and which are learned through visual experience. Infant work has identified common motion as an important grouping cue early in development. Our own work in Project Prakash (Nature, 2006, 441, 271–272) has shown that late-onset vision patients, almost immediately after sight recovery, are able to make use of motion, but not other grouping cues (such as continuation, junction parsing, and color) to parse visual imagery. Building upon these results, here we explore the process by which learning heuristics might be acquired, and, specifically, the role of motion in this process. We conducted computational analyses to examine the relative efficacies of motion and color cues for learning grouping heuristics. Working with natural visual sequences, we find that common motion cues, as compared to common color or luminance cues, are more effective for learning at least one important Gestalt heuristic - grouping via contour continuation. Motion-mediated learning, we find, can proceed with fewer training samples and is effective over larger spatial distances. We infer that motion cues, more than others, might embody reliable statistical properties in the natural world. To complement this computational result, we are conducting empirical studies with normal adults to examine whether the privileged status of motion as a grouping mechanism is evident in the mature visual system. By pitting motion cues against other binding cues in behavioral grouping tasks, we are able to titrate the contribution of different cues and arrange them in a tentative hierarchy based on their grouping efficacy.