抄録

◎Ruihan Bao・Kyota Higa・Kota Iwamoto（NEC）

In this paper, we propose a fast and robust geometric verification method for local descriptor based image matching, serving as an alternative to RANdom SAmple Consensus (RANSAC) based method. Instead of iteratively solving for transformation models and checking the maximum number of keypoint matches consistent with the models, we cluster the parameters of similarity transformation models computed directly from keypoint matches using keypoint scale, orientation and coordinates. The resultant cluster size is used as a criterion for geometric consistency between two images. We evaluated the proposed method on dataset of images containing planar and 3D objects taken from different viewpoints, and experiment results show that our method not only reduces the processing time but also achieves higher matching accuracy compared with conventional RANSAC based methods.