Improving RANSAC filtering with matching similarity of local features

We improved RANSAC filtering with similarity between matching pairs of local features. As knowns, local features, such as SIFT, use descriptor for matching between different feature set. The method of distance ratio of nearest neighbors are employed to judge whether a point pair is an acceptable matching pair or not. Our proposed similarity is determined by distance ratio of nearest neighbors among matching pairs. An ordered queue of similarities corresponding to matching pairs is built up. The matching pairs in front part of the similarity queue are utilized in a linear-order combination to compute the transformation matrix. Experiments show that the proposed method can find more inliers than original RANSAC and can generate an accurate transformation matrix in a few iterations. The advancement of the proposed method is due to the priori knowledge of similarity of matching pairs.