Abstract
Research suggests that, due to capacity limitations, the visual system pools information over sizable regions, which grow linearly with eccentricity. In many artificial experiments, this causes pooling over uninformative "flankers", leading to crowding. However, under natural circumstances, objects are typically surrounded by informative context. In normal viewing, how does the harmful effect of pooling over a large, potentially complex region (i.e. crowding) trade off against the beneficial effect of additional context? We conducted a recognition experiment in which we varied the size of the contextual region surrounding the object. 656 objects were randomly selected from a fully annotated picture database (SUN 2012). Objects were presented at 10 degrees from the fovea, and subtended 4 degrees visual angle. In one condition, the objects appeared isolated from the background. Otherwise, the objects appeared within a circular cropping of the original picture, with radius varying from 1 (object size) to 5 times the object size. In addition, we examined accuracy identifying the object from the context alone (largest window size into the scene, object occluded by a patch the size of the smallest window). Recognition performance was 36% for the cut out objects, then increased monotonically from 45% to 71% with increasing window size, showing no detrimental effect of increasing the surround to include the typical "crowding zone". Performance with context alone was 29% correct. These results confirm that there are object recognition benefits to pooling information over a large region. The visual system, faced with capacity limitations, has made a reasonable compromise. On average, for real world identification, contextual information more than makes up for the loss of information underlying crowding.
Meeting abstract presented at VSS 2014