Max-Matching Context Kernel Design for SIFT Feature

In object matching and recognition it is useful to represent an image with the sets of local features such as SIFT etc. However this representation poses a challenge to the popular SVM machine learning method, since it needs ordered and fixlength data. To solve this problem, we focus in this paper on a Max-matching context kernel, which computes Max-matching points based on the angle between two points of sets and sums its context energy to Max-matching point energy. We prove Max-matching context kernel yields a Mercer kernel. We demonstrate our algorithm on Caltech 101's object recognition, which shows that the proposed method is accurate and significantly more efficient than current other approaches.