Propagation for feature matching using triangular constraints

This paper presents a novel Matching Propagation Framework for addressing the problem of finding better matching pairs between each two images, which is one of the most fundamental tasks in computer vision and pattern recognition. We first select initial seed points by original matching method like SIFT, and then use T-CM to explore more seed points. Finally, a triangle constraint based quasi-dense algorithm is adopted to propagate better matches around seed points. The experimental evaluation shows that our method can get a more precise matching result than classical quasi-dense algorithm. And the 3D reconstruction of the scene from our method has a good visual effect. Both experiments demonstrate the robust performance of our method.