Two-step clustering of SIFT keypoints and relaxation based matching of clusters
Sunmin Lee, Yong Cheol Kim · 2015
Modern buildings in urban cities have repetitive structure such as windows or wall patterns. SIFT keypoints extracted from those buildings are so similar to each other that finding the most similar match between repetitive keypoints in building images has inherent ambiguity. Conventional similarity-based matching is highly prone to error. We propose a relaxation based matching of clusters of SIFT keypoints. First, the proposed two step mean shift clustering successfully groups keypoints in accordance with their structural and locational homogeneity. Then, relaxation based matching reliably matches clusters, by preserving the structural consistency among matched clusters over matched frames. We tested the proposed methods on several images of buildings and achieved significant improvement both in the precision rate and in the recall rate.