Feb. 17, 11:45, Wei Ji Ma. Optimality and Probabilistic Computation in Visual Categorization

Anatoliy Kats, Rutgers

Medial part abstraction and grouping from images

Medial axis-based representations represent a powerful class of symmetry-based shape decomposition that supports effective shape categorization. However, their applicability to real scenes has been severely limited due to their restrictive assumption that the Object’s silhouette is available. We present a novel framework for recovering a set of symmetric parts from a cluttered scene, and grouping them to form an approximation to an object’s medial axis transform. Starting with a set of extracted image contours, one of the key contributions is a transformation of the symmetry-based perceptual grouping problem from the space of image contours to the dual space of medial contours. Given the graph of hypothesized medial axis fragments, we introduce a weight function that captures the nonaccidental relations between pairs of these fragments, and use it to group together medial fragments that belong to the same part, or medial branch. Defining a graph over the recovered parts, we introduce a second weight function that drives the grouping of parts into objects, effectively assembling the medial branches to yield the final medial axis approximation. The ability to recover a medial axis approximation of an object from a cluttered image will allow a rich community of skeleton-based recognition algorithms access to real-world imaging conditions.