Event Details

Visual flows are a rich source of information for qualitativescene properties, in the manner that shading, classically,conveys information about surfaces. However,local measurements of flows are noisy and incomplete, making global inferencesdifficult. Gestalt good continuation provides a framework for extending local measurements,which we develop from a differential-geometric perspective.The result is a flow relaxation network modelled on the properties of primate visual cortex.We concentrate on static patterns, including texture, shading, color,and stereo, and demonstrate that robust computationof visual flows predicts the physiological distribution of long-range horizontalconnections. Combinations of flows provide new insight into visualinferences, such as border ownership and shadow detection.