To this end, we created sets of images that parametrically varied between objects and textures, and recorded responses to these images from populations of single units in macaque mid-level visual cortex (V4). We used a common texture synthesis algorithm (Portilla-Simoncelli) that preserves statistics of the original image while discarding information about scene and shape. We adapted the algorithm to parametrically vary the size of the regions in which statistics were measured (the “scrambling regions”), resulting in sets of images, matched in texture, that smoothly varied from photographic to fully scrambled.

V4 responses were sensitive to shape. Although neurons responded equally strongly to photographic images and their scrambled counterparts, classification analyses revealed that populations were 50% better at discriminating pairs of photographic images than pairs of scrambled images. However, V4’s shape sensitivity was highly sensitive to small distortions in photographic structure. Classifiers could discriminate between photographic and scrambled images, but were impaired when discriminating lightly scrambled images from fully scrambled images. This sensitivity was not due to low-level properties of the image: standard image similarity metrics did not show this property. It could, however, be replicated by a CNN-based metric trained to predict human judgments of image degradation (DISTS).

These measurements reveal that mid-level processing in primate vision relies significantly on shape cues, and establish baseline metrics for future models of visual processing.