MGCNet: Multi-granularity consensus network for remote sensing image correspondence pruning

Correspondence pruning aims to remove false correspondences (outliers) from an initial putative correspondence set. This process holds significant importance and serves as a fundamental step in various applications within the fields of remote sensing and photogrammetry. The presence of noise, illumination changes, and small overlaps in remote sensing images frequently result in a substantial number of outliers within the initial set, thereby rendering the correspondence pruning notably challenging. Although the spatial consensus of correspondences has been widely used to determine the correctness of each correspondence, achieving uniform consensus can be challenging due to the uneven distribution of correspondences. Existing works have mainly focused on either local or global consensus, with a very small perspective or large perspective, respectively. They often ignore the moderate perspective between local and global consensus, called group consensus, which serves as a buffering organization from local to global consensus, hence leading to insufficient correspondence consensus aggregation. To address this issue, we propose a multi-granularity consensus network (MGCNet) to achieve consensus across regions of different scales, which leverages local, group, and global consensus to accomplish robust and accurate correspondence pruning. Specifically, we introduce a GroupGCN module that randomly divides the initial correspondences into several groups and then focuses on group consensus and acts as a buffer organization from local to global consensus. Additionally, we propose a Multi-level Local Feature Aggregation Module that adapts to the size of the local neighborhood to capture local consensus and a Multi-order Global Feature Module to enhance the richness of the global consensus. Experimental results demonstrate that MGCNet outperforms state-of-the-art methods on various tasks, highlighting the superiority and great generalization of our method. In particular, we achieve 3.95% and 8.5% mAP$5°$ improvement without RANSAC on the YFCC100M dataset in known and unknown scenes for pose estimation, compared to the second-best models (MSA-LFC and CLNet). Source code: https://github.com/1211193023/MGCNet.