MSSF: A Novel Mutual Structure Shift Feature for Removing Incorrect Keypoint Correspondences between Images

Removing incorrect keypoint correspondences between two images is a fundamental yet challenging task in computer vision. A popular pipeline first computes feature vector for each correspondence and then trains binary classifier using these features. In this paper, we propose novel robust to better fulfill the above task. The basic observation that relative order of neighboring points around correct match should be consistent from one view another, while it may change lot an match. To end, designed measure bidirectional ranking difference neighbors reference correspondence. reduce negative effect neighborhood when computing feature, combine spatially nearest with geometrically “good” neighbors. We also design iterative neighbor weighting strategy, which considers both goodness correctness correspondence, enhance suppress correspondences. As encodes structure information them, name proposed Mutual Structure Shift Feature (MSSF). Finally, use features train random forest supervised manner. Extensive experiments on raw matching quality downstream tasks are conducted verify performance method.