Efficient Object Recognition Using Sampling of Keypoint Triples and Keygraph Structure

We present an object matching method that employs matches of local graphs of keypoints, called keygraphs, instead of simple keypoint matches. For a keygraph match to be valid, vertex (keypoint) descriptors must be similar and both keygraphs must satisfy structural properties concerning keypoints orientation, scale, relative position and cyclic ordering; as a result, the large majority of initial incorrect keypoint matches is correctly filtered out. We introduce a novel approach to sample keypoint triples (i.e. keygraphs) in a query image, based on complementary Delaunay triangulations; this generates a linear number of triples with relation to the number of keypoints. Query keygraphs are then matched against the indexed model keypoints; each established keygraph match is used to evaluate a candidate pose (an affine transformation). The proposed method has been evaluated for object recognition and pose estimation, achieving a better performance in comparison to state-of-the-art methods.