Abstract
Humans are capable of perceiving and recognizing objects even when part of the object is obscured from view. The perception of entire objects despite their occlusion is called amodal completion. Several studies have shown that responses in high-level visual areas are invariant to occlusion (Hulme & Zeki, 2007; Weigelt et al., 2007), suggesting that amodal completion has occurred by this stage of the visual hierarchy. However, there is less evidence for amodal completion in early visual cortex. Here we present neuroimaging findings that suggest quite similar levels of object completion occur in early and late visual cortex. We presented 8 different naturalistic object images, either in full view or with one of two complementary occluders. Each occluder consisted of 8 horizontal black bars obscuring 50% of the image with the two occluders covering entirely non-overlapping areas of the image. Thus, there was no shared image information across two differently occluded images of the same object exemplar. Nevertheless, both lower visual areas (V1, V2, V3) and higher visual areas (IT) showed similar patterns of response when the same object was presented with either occluder. In fact, for every region, the strength of the similarity was not statistically different to that of two identical occluded images, suggesting that early visual areas attained full invariance to the different occluder positions. To explore whether these results could be explained by feedforward or recurrent mechanisms (Svanera et al., 2021), we examined responses to these images in a feed-forward deep neural network trained to classify objects (Krizhevsky et al., 2012). We found that responses in early layers were more occluder dependent while responses in later layers were more occluder invariant. This indicates top-down recurrent processing as a plausible explanation for the occluder invariance observed across the entire visual pathway.