Abstract
The ventral visual stream has extensive bidirectional connectivity that has been suggested to promote recurrent processing – what function might this serve in the context of high-level visual processes like object recognition? Previous computational modeling work from our lab has demonstrated that the bidirectional architecture of the ventral visual stream supports a highly robust object recognition system. Specifically, our bidirectional model successfully categorized visual objects from 100 different real-world categories, even under high levels of occlusion, with up to twice the accuracy of a comparable feedforward-only model. Here, we expand on this work by experimentally exploring the recurrent processing dynamics that give rise to the visual system's robustness to occlusion during object recognition. Human subjects categorized images of occluded object stimuli that were followed by a mask on some trials. As predicted by our model, we found a significant interaction between the mask (present or absent) and occlusion (low or moderate) such that recognition performance was differentially impaired when a moderately occluded stimulus was masked compared to a relatively unoccluded one. The model provided a close quantitative fit to subjects' data on the identical stimuli and task parameters as used in the experiment. The model indicated that the mask interfered specifically with the extensive recurrent processing required to resolve the ambiguity present in moderately occluded stimuli, whereas low occlusion inputs could be rapidly resolved and thus were relatively unaffected by the mask. Together, the results of this experiment and the accompanying modeling work provide support for the view that object recognition is a highly dynamic process that depends on the bidirectional architecture of the ventral visual stream.
NSF SBE0542013, ONR N00014-07-1-0651, ONR N00014-10-1-0177, iARPA/ARL W911NF-10-C-0064.