2D Shape Classification and Retrieval

Abstract

We present a novel correspondence-based technique for efficient shape classification and retrieval. Shape boundaries are described by a set of (ad hoc) equally spaced points – avoiding the need to extract “landmark points”. By formulating the correspondence problem in terms of a simple generative model, we are able to efficiently compute matches that incorporate scale, translation, rotation and reflection invariance. A hierarchical scheme with likelihood cut-off provides additional speed-up. In contrast to many shape descriptors, the concept of a mean (prototype) shape follows naturally in this setting. This enables model based classification, greatly reducing the cost of the testing phase. Equal spacing of points can be defined in terms of either perimeter distance or radial angle. It is shown that combining the two leads to improved classification/retrieval performance.