A frame-based probabilistic local verification method for robust correspondence

Establishing reliable feature correspondence between two sets of features is a fundamental task in image processing. In this paper, we propose a novel probabilistic local verification method to reject false feature matches. We exploit the local affine frame to calculate the re-projection error, and develop a novel probabilistic model to estimate the correspondence confidence according to the error. The correspondence confidence is evaluated by calculating the posterior probability based on a two-layer mixture model. The key parameters of the proposed method can be adaptively estimated by alternatively maximizing and updating a second lower bound function. We also suggest that the adjacent inlier neighbors are good neighbors and thereby proposing a confidence-distance-ratio strategy to balance the inlier confidence and spatial distance. Our method mostly outperforms other state-of-the-art methods by over ten percentage points in the success rate of UAV localization tasks, and by over six percentage points in the F-measure on multiple public test datasets. The code is available at https://github.com/shenliang16/Iterative-Probabilistic-local-Verification.