A viewpoint invariant signature descriptor for curved shape recognition

Abstract

Shape descriptor plays important role in model based subject (object or motion) recognition. A difficulty is that the projected views of the same subject may vary with respect to the changes of viewpoint (camera pose). Therefore, viewpoint invariant descriptors are desired. To this end, many geometric invariants based projective invariants have been studied. However, it is rather hard to use most geometric invariants to represent nature subjects as the geometric invariants require the subject to have certain specific geometric configuration (combination of relation-constrained points, lines or planes). In addition, some reported descriptors based on global features suffer from occlusion. In this paper, focusing on the local curve features, we propose a viewpoint invariant signature descriptor for curved shape recognition. As the descriptive geometry element is curve-oriented, the signature is capable of describing complex subjects. Further, occlusion can be well resolved due to the utilization of the local differential invariants. More importantly, to avoid the noise-sensitive high order derivatives, a reliable approximate signature is implemented. The nonlinear inter-signature matching metric is also customized to perform shape recognition. The conducted experiments verified the signature's effectiveness and its viewpoint invariant.