OAMatcher: An overlapping areas-based network with label credibility for robust and accurate feature matching

Abstract

Local feature matching involves establishing accurate pixel-wise correspondences between an image pair, which is a critical component in several visual applications (e.g., visual localization). Recently, detector-free techniques have realized excellent performance in this task. However, existing methods tend to focus on the entire image without prioritizing overlapping regions, resulting in undesirable interference from non-overlapping areas during the descriptors enhancement process. Moreover, these approaches neglect unreliable ground-truth matching labels triggered by measurement noise in datasets, leading to sub-optimal network optimization. In this study, we develop a novel overlapping areas-based network OAMatcher to resolve these issues. For the first issue, OAMatcher employs an overlapping regions perception block (ORPB) that captures the overlapping areas of image pairs to filter out plentiful mismatches and circumvent interference from non-overlapping regions during descriptors enhancement process. Specifically, the ORPB first enhances the descriptors of all keypoints to mimic the human behaviour of scrutinizing entire images back and forth at the start of feature matching. Subsequently, the ORPB introduces an overlapping regions extraction block (OREB) that captures the keypoints within overlapping zones to mimic the humans behaviour of shifting the focus from the whole images to co-visible areas. After OREB, ORPB performs descriptors enhancement exclusively among the keypoints within these co-visible regions, ensuring minimal disturbances from non-overlapping areas. In addition, the ORPB confines the predicted matches strictly to co-visible regions, thus efficiently filtering out a significant number of mismatches in non-overlapping zones. For the second issue, OAMatcher proposes a labels weighting algorithm (LWA) that predicts the label credibility for ground-truth matching labels. LWA assigns low credibility to unreliable labels and utilizes the credibility to weight loss, effectively diminishing the influence of unreliable labels. Extensive experiments show that OAMatcher delivers excellent results for homography estimation, pose estimation, and visual localization tasks.