Rejecting Mismatches between Fish-Eye Camera Images by RVM

Establishing reliable correspondence points is a fundamental problem in computer vision. In this work, we studied mismatch-rejecting between two fish-eye camera image by RVM learnings. The fundamental idea that, for given two fish-eye images of a scene, the corresponding points constitute a manifold in joint-image space R⁴, and outliers can be detected by checking whether they are consistent with the upward views of the manifold. Experiments on real image pairs demonstrate the excellent performance and feasibility of our proposed method.