Combining hierarchical segmentation and shape context based recognition

In this paper, we integrate image segmentation and object recognition into one unified process. The process starts from a seed region produced by the initial segmentation. Then it is evolved through alternating shape context based recognition and smaller scale image segmentation, until an object is found. In recognition, the thin plate spline (TPS) transformation is employed to locate the difference between the seed region and the defined model, and guide the subsequent evolution process. Experimental results demonstrate the advantages achieved by the combination of hierarchical segmentation and multi-scale shape contexts based recognition.