| This thesis explores two key aspects of the problem of object recognition: image segmentation and shape matching. We approach the first problems in the Normalized Cut framework, wherein we focus on the problem of integrating the cues of contour and texture. For this purpose, we introduce an operationalization of Julesz' concept of the “texton,” analogous to a phoneme in speech recognition in terms of linear filter outputs. This representation, together with a novel gating mechanism for they cue of brightness boundaries, allows us to segment images containing both textured regions and regions with smooth shading.; Next we address the problem of matching shapes—a process that assumes that a segmentation is available. For this purpose we introduce a new descriptor called the “shape context.” Using a sampled point set representation of shape, the shape context captures the distribution of shape points relative to a reference point and thus offers a globally discriminative characterization for each shape point. When incorporated into an iterative framework including a thin plate spline coordinate transformation estimation step, shape contexts provide us with robust correspondences and alignment between shapes.; Finally, we discuss the design of an end-to-end system that combines the above approaches to segmentation and shape matching to make a complete object recognition system. |
