Two-dimensional shape matching by shape distributions

Two-dimensional shape matching is a fundamental problem in pattern recognition and computer vision. A challenging aspect of this problem is to find a distinctive shape descriptor which is able to handle common geometric transformations, occlusions and deformations. In this paper, we present a novel and distinctive shape descriptor based on shape distributions. The key concept of our method is that based on the defined three distance functions, the shape descriptor is built by a combination of three shape distributions. The proposed shape descriptor is not only invariant to rotation, translation and scale but also insensitive to shape occlusion and deformation. The shape similarity is measured as the weighted distance sum of three distributions. Experimental results on different shape databases show that our method outperforms several well-known algorithms.