Document Type

Technical Report

Publication Date


Technical Report Number



In this paper, we propose a new class of kernels for object recognition based on local image feature representations. Formal proofs are given to show that these kernels satisfy the Mercer condition and reflect similarities between sets of local features. In addition, multiple types of local features and semilocal constraints are incorporated to reduce mismatches between local features, thus further improve the classification performance. Experimental results of SVM classifiers coupled with the proposed kernels are reported on ecognition tasks with the standard COIL-100 database and compared with existing methods. The proposed kernels achieved satisfactory performance and were robust to changes in object configurations and image degradations.


Submitted to CVPR 2005.