Conference on Cognitive Computational Neuroscience

Humans are able to perceive objects as complete wholes even when the retinal image is partial or fragmented due to occlusion by other objects, a phenomenon known as amodal completion. The literature contains mixed results regarding the extent to which amodal completion occurs across different levels of the visual hierarchy; moreover, the functional role of attention in completion has not been fully explored. We presented 8 different naturalistic object images with one of two complementary occluders which covered non-overlapping halves of the image. Despite these stimulus differences, fMRI response patterns to differently occluded images of the same object were statistically indistinguishable in both early and high-level visual cortex, suggesting that all visual areas attained full completion of the object representation. We also observed the same pattern of results when attention was directed away from the object, suggesting that attention is not necessary for amodal completion. Finally, we measured unit responses to occluded images in a recurrent deep neural network designed to mirror the ventral visual hierarchy (CORnet-S). We found that this network differed from humans in that object completion emerged gradually across layers. Taken together, these results suggest that humans and recurrent networks may attain occlusion robustness through different mechanisms.