Japanese / English

Detail of Publication

Text Language Japanese
Authors Kazuto NOGUCHI, Koichi KISE, and Masakazu IWAMURA
Title Relation Between the Object Recognition Rate and the Matching Accuracy of Local Descriptors
Journal Proceedings of MIRU 2008
Presentation number IS4-4
Pages p.504
Reviewed or not Not reviewed
Month & Year July 2008
Abstract A simple method of object recognition with local descriptors is voting by matching local descriptors with nearest neighbor (NN) search. Because the number of local descriptors is so large, a key technology for this approach is to reduce the computational cost of matching. It is known that the cost of NN search can be drastically cut by approximation, though it poses another question of how accurate the matching of local descriptors should be for accurate object recognition. The major contribution of this report is twofold. First, we experimentally show that it is not necessary for matching to be accurate. For example, an object recognition rate of 97 % with 100,000 object models is achieved by the matching accuracy of only 14%. Secondly, we propose a model of recognition by voting based on a simple binomial distribution, which agrees with experimental results.
Back to list