Research

Detection and recognition of objects in a natural scene are fundamental components of vision, yet the underlying cortical computations are not understood. Planar shape – the silhouette boundary contour of a solid body – carries rich information about the nature of objects and scenes. In the human vision system, it is well-known that object properties like identity and pose are encoded by neurons of inferior temporal cortex (IT), and oriented contour elements are encoded by neurons of primary visual cortex (V1). However, it remains unclear how intermediate ventral areas V2 and V4 integrate signals from V1 into the representations of object shape in IT that guide our behaviour in natural environments.

Previous work has described aspects of intermediate shape representations in primate cortex (e.g., border ownership in V2 and boundary curvature in V4) and artificial neural networks (e.g., intermediate layers of machine vision systems trained for object recognition).  However, these results are typically limited by using artificial shape stimuli without natural object features or natural images in which planar shape is not controlled.

To overcome these limitations, we utilize a generative model for shape in which local features (i.e., curvature) are learned from natural objects. This method overcomes the challenges of satisfying non-intersection and closure constraints and creates planar shapes that match the statistics of natural object boundaries.  My independent research program as Assistant Professor at the University of Regina aims to combine these interdisciplinary insights with a comprehensive computational approach to model the neural computations that underlie natural object perception.