A novel shape descriptor based on von Mises distributions

Since most local descriptors of shape are not scale invariant, we usually make the line drawings or object contours in an image database the same size through scale normalization, before applying shape recognition procedures. Drawings produced by scale normalization are suitable for such descriptors if the whole of the original drawings are similar in shape. They are, however, not suitable if parts of each drawing are drawn using a different scale. In this paper, we propose a novel scale invariant descriptor that does not require scale normalization. The experimental results on shape matching and retrieval show the effectiveness of our descriptor, compared to several conventional descriptors.