Image Retrieval Using Shape Contexts

In this paper we have used shape models that are computationally fast and invariant to basic transformations like translation, rotation and scaling. This work drives shape detection using a feature called shape context. We demonstrate that shape context can be used to quickly prune similar shapes. Shape context describes all boundary points of a shape with respect to any single boundary point. The shape context at a reference point captures the distribution of the remaining points relative to centroid, thus offering a globally discriminative characterization. Corresponding points on two similar shapes will have similar shape contexts, enabling us to solve for correspondences as an optimal assignment problem. Thus it is descriptive of the shape of the object. Object recognition can be achieved by matching this feature with a priori knowledge of the shape context of the boundary points of the object.