Abstract
Blur is a common and informative cue of naturalistic visual scenes. For example, cast shadows have blurry boundaries, as do objects outside the focal plane, and surface features of 3D objects may be associated with shading blur. Interestingly, while psychophysical studies have long demonstrated the importance of detecting and encoding blur for scene segmentation and perception of depth and 3D structure, the underlying neural mechanisms have yet to be discovered. To investigate this we record single-unit activity from area V4 in two awake fixating Macaca mulatta in response to shape stimuli exhibiting blurred boundaries. Specifically, after classifying shape selectivity of single neurons, preferred and non-preferred shapes are presented under multiple levels of Gaussian blur. Surprisingly, our data reveals a population of V4 neurons which are tuned for intermediate levels of boundary blur, demonstrating, for the very first time, blur selectivity anywhere in primate visual cortex. After performing a series of control experiments our results reveal a sophisticated neural computation within V4, with responses being enhanced by the removal of high spatial frequency content; this effect is not explained by confounding factors of stimulus size, curvature, or contrast. We interpret our findings in the context of computational studies that argue for shape and blur as forming a sufficient representation of naturalistic images. A simple descriptive model is proposed to explain observed data, wherein blur selectivity modulates the gain of shape-selective responses, supporting the hypothesis that shape and blur are fundamental features of a sufficient neural code for natural image representation within the ventral pathway. More generally, we believe that our findings will shift paradigms surrounding area V4's role in visual processing: as opposed to computations of object recognition alone, our results suggest that V4 also provides the neural substrate underlying processes of scene segmentation and understanding.
Meeting abstract presented at VSS 2017