Application of Shape Context and Belief Propagation in Removing SIFT Mismatches

A matching method based on belief propagation and shape context is proposed to screen out the mismatches of SIFT algorithm. The proposed method includes four steps. 1) The local shape context of each SIFT feature point is calculated within each feature’s local neighborhood. The size and orientation of each feature’s neighborhood is determined by each feature’s scale and primary orientation. 2) Each feature is connected with its three nearest neighbor features to constitute the skeleton frame of the belief propagation net. 3) The evidence function is designed by using the descriptor distance and local shape context distance between each pair of initial SIFT correspondent features, and the consistent function is constructed by using the spatial geometric information between each pair of correspondent features and their nearest neighbors. 4) Each feature’s belief and the massages passing to its three nearest neighbors are calculated iteratively until it converges, and the converged beliefs of the features are used to screen out SIFT mismatches. The simulation was conducted and compared with RANSAC, GTM and BP-SIFT with real captured images and the images of Oxford Visual Geometry Group. The simulating results certify that the proposed method generally performs better than RANSAC, GTM and BP-SIFT in recall rate, precision rate, loss rate and efficiency.