Abstract
Correspondence selection between two groups of feature points aims tocorrectly recognize the consistent matches (inliers) from the initial noisymatches. The selection is generally challenging since the initial matches aregenerally extremely unbalanced, where outliers can easily dominate. Moreover,random distributions of outliers lead to the limited robustness of previousworks when applied to different scenarios. To address this issue, we propose todenoise correspondences with a local-to-global consensus learning framework torobustly identify correspondence. A novel "pruning" block is introduced todistill reliable candidates from initial matches according to their consensusscores estimated by dynamic graphs from local to global regions. The proposedcorrespondence denoising is progressively achieved by stacking multiple pruningblocks sequentially. Our method outperforms state-of-the-arts on robust linefitting, wide-baseline image matching and image localization benchmarks bynoticeable margins and shows promising generalization capability on differentdistributions of initial matches.