Joint Template Matching by Candidate Set Voting

To jointly match multiply images for one specific object has more and more applications in practice. In this paper, given a small annotated subset as templates, we propose to jointly match a series target images by a novel idea: selecting the initial matching results by the bidirectional voting between template set and candidate set. In our method, firstly, the forward pairwise matching from template set to target set is performed. These matching results are denoted as candidates. When multiply results appear for one target, that means the matching results are inconsistent. Then, a novel screening method is proposed by the backward matching and voting, until all results reach consistent. Experiments conducted on various moving objects in the wild demonstrate the proposed method can overcome the matching difficulty by a single template and improve the matching accuracy under complex conditions.