Part-Based Shape Recognition Using Gradient Vector Field Histograms

摘要

The gradient vector field generated from the boundary of a shape describes the regional interaction between the shape boundaries and can therefore be exploited to provide rich and robust shape description. We present a novel part-based shape representation that describes a shape using a set of gradient vector field histograms derived at salient points within the shape. Peaks and ridges derived from the local disparity in the vector field provides a means of locating these salient points called shape axes, from where polar sampling of the vector field is then used to build scale and rotational invariant histograms of the vectors' orientation. A multi-resolution pyramidal framework is proposed for generating the gradient vector field and extracting the shape axes. Results from shape recognition experiments show that the proposed shape descriptor is invariant to similarity transform, robust under boundary distortion and occlusion. This part-based descriptor also supports partial matching and articulation.