1.5 million subspaces of a local feature space for 3D object recognition

We propose 3D object recognition methods whose characteristic point is the use of a large number of subspaces (1.5 million) generated from a billion of local features for the recognition of 1002 objects. In order to match query local features to a lot of subspaces, a simple approximate nearest neighbor search is utilized. Based on this approximation the proposed three methods are as follows: a method with an ordinary subspace method (match each query local feature to subspaces), a method of two-step matching which employs two versions of subspaces for efficient matching, and a method with a mutual subspace method (both queries and models are represented as subspaces). These methods are compared with two baselines with and without subspaces. From experimental results, we have confirmed the advantage of proposed methods in both accuracy and efficiency. In particular, the mutual subspace method achieves 95% accuracy with processing time of 3.5 sec./query, which improves the accuracy of a baseline without subspaces about 60%. As compared to the subspace method without approximate matching, the mutual subspace method is more than 240 times faster.