Japanese / English

Detail of Publication

Text Language Japanese
Authors Takayuki HONDO, Koichi KISE
Title Performance Evaluation of Local Features for Large-scale Image Recognition
Journal Proceedings of MIRU 2008
Presentation number IS5-6
Pages p.550
Reviewed or not Not reviewed
Month & Year July 2008
Abstract Various interest points/regions detectors and local descriptors have been proposed for image recognition. However, it is still unclear which interest points/regions detector or local descriptor gives the best performance for large-scale image recognition of object instances(specific objects). In this report, we present the results of comparison about performance of image recognition with several interest points/regions detectors and local descriptors using a 10,000 image-database. In our experiment, we confirmed that it was the best to employ Hessian-Affine Region as the interest points/regions detector and GLOH/Shape Context as the local descriptor. And PCA-SIFT, providing the functions of both the detector and the descriptor, gives the best performance as the combinational method. In addition, we confirmed that the tendency of misrecognition is different even if the same local descriptor is employed, in case of employing interest regions computed with another detector, furthermore, combination of detectors gives higher accuracy.
Back to list