Abstract
Image segmentation – the process by which scenes are segmented into component objects – is a fundamental aspect of vision and a cornerstone of scene understanding; its neural basis, however, is largely unknown. Partial occlusions pose a special challenge to segmentation because, unlike non-overlapping stimuli, they require the parsing of overlapping contours and regions and/or the grouping of noncontiguous regions. To begin to understand how partially occluded stimuli are segmented in the primate brain, we studied the responses of single neurons in IT cortex to shape stimuli subjected to increasing levels of occlusion. We asked whether IT responses are consistent with a segmented representation whereby responses of each neuron are dictated by either the occluded or the occluding stimulus, but not both. We recorded from 43 well-isolated, single IT neurons as animals were engaged on a sequential shape discrimination task. On each trial, two stimuli were presented in sequence and the animal had to report whether the stimuli were the same or different with a rightward or leftward saccade, respectively. The second stimulus in the sequence was occluded with randomly positioned dots; occlusion levels were titrated by varying occluding dot width. Some neurons (11/43, 26%) showed strong responses to unoccluded stimuli and responses gradually declined with increasing levels of occlusion. These unoccluded-preferred neurons showed shape-selective responses to occluded stimuli. These neurons behaved quite like those in IT cortex during passive fixation (Kovacs et al., 1995) and their responses were consistent with a encoding of the identity of the occluded shape. Many others (21/43, 49%), however, showed weak responses to unoccluded stimuli and stronger responses under occlusion. Taken together, our results support the idea that IT neurons encode segmented components of the image, with one sub-group encoding the occluded stimulus and other encoding the occluders.
Meeting abstract presented at VSS 2017