Clustering SIFT: An Efficient Way to Achieve Flipping Invariance

The design and improvement of feature descriptor is always a prevalent research point. A strong design of feature descriptor can be applied to many industrial fields. However, in many successful designs, there are still have some shortcomings that need to be improved like the flipping operation in image. Although there are many techniques to deal with this problem recently, they still exist some uncertainty factors because they do not compare with different images which have the same but flipped object in their experiments. In this paper, we design an unsupervised learning structure algorithm based on SIFT to achieve flipping invariant and change object to get as many correct matches as possible. Then we will use this algorithm to match in UKBench dataset and indicate that this approach not only solves the problem of flipping operation but also ensures high matching accuracy. Compared with previous other algorithms, our algorithm is extremely simple to implement and does not destroy the structure of SIFT.