Match propagation starting from a single initial correspondence under wide baseline views

It is important for mobile robot to match features from scenes under wide baseline views. Wide baseline matching is one of the difficult problems in feature matching because large variation between the image appearances occurs due to illumination and viewpoints, especially on non-planar and less-textured scenes. In this paper we propose a method of feature matching under large separate views, which is based on the multiple hypotheses about the local transformation and multiple routes propagation. Given a single initial correspondence and the local affine transformation between the neighborhoods around them, the proposed method can supply more stable and accurate matches under wide baseline views. The matches can be propagated adaptively into the regions where the local transformations differ from the initial one. The experiments with real data show that the proposed method achieves the state-of-the art performance among match propagation algorithm.