Invariant multi-scale descriptor for shape representation, matching and retrieval

Shape matching and retrieval have been some of the fundamental topics in computer vision. Object shape is a meaningful and informative cue in object recognition, where an effective shape descriptor plays an important role. To capture the invariant features of both local shape details and visual parts, we propose a novel invariant multi-scale descriptor for shape matching and retrieval. In this work, we define three types of invariants to capture the shape features from different aspects. Each type of the invariants is used in multiple scales from a local range to a semi-global part. An adaptive discrete contour evolution method is also proposed to extract the salient feature points of a shape contour for compact representation. Shape matching is performed using the dynamic programming algorithm. The proposed method is invariant to rotation, scale variation, intra-class variation, articulated deformation and partial occlusion. Our method is robust to noise as well. To validate the invariance and robustness of our proposed method, we perform experiments on multiple benchmark datasets, including MPEG-7, Kimia and articulated shape datasets. The competitive results indicate the effectiveness of our proposed method for shape matching and retrieval.