Image matching via higher order structures

Image matching is an important and fundamental problem in computer vision. When human matches two objects in images, not only local feature, (i.e. colour similarity and etc.), but also higher order structures (i.e. shape similarity and etc.) are used. In this paper, a novel higher order score function for image matching is proposed, and then the score function are optimised via efficient belief propagation algorithm. Experiments show that the proposed image matching method are robust to noise, JPEG compression and etc. The proposed algorithm also handles large scale problems that previous algorithms fails to handle.