Shape Representation and Recognition in High Dimensional Feature Space

We present a novel approach for shape representation using support vector domain description (SVDD). The shape contour of an object is mapped to spherical surface in high dimensional feature space, where similar contours with different size look like concentric spheres in the learned feature/kernel space. Shape matching in cluttered image, therefore, can be seen as finding the corresponding super spherical surface after the nonlinear transformation. The shape model, which consists of several support vectors (SVs), is partially invariant to contour scaling and rotation. A coarse-to-fine strategy is then adopted for shape matching in the forward matching process. Experimental results show that the approach is robust and suitable for shape-based image retrieval. In addition, our approach enables the user to fast construct the template shape by a small quantity of points which represent the shape roughly. The precise shape contour is not needed. It can greatly improve the human-machine interaction friendliness of practical image retrieval systems. Though our analysis is based on 2D shapes, this idea can be easily extended to the 3D shapes.