E2LSH based multiple kernel approach for object detection

Multiple kernel learning (MKL) methods is widely used in object detection. The conventional MKL methods employ a linear and stationary kernel combination format which cannot accurately describe the distributions of complex data. This paper proposes an E2LSH based clustering algorithm which combines the advantages of nonlinear multiple kernel combination methods-E2LSH-MKL. E2LSH-MKL is a nonlinear and nonstationary multiple kernel learning method. This method utilizes the Hadamard product to realize nonlinear combination of multiple different kernels in order to make full use of information generated from the nonlinear interaction of different kernels. Besides, the method employs E2LSH-based clustering algorithm to group images into subsets, then assigns cluster-related kernel weights according to relative contributions of different kernels on each image subset to realize nonstationary weighting of multiple kernels to improve learning performance. Finally, E2LSH-MKL is applied to object detection. Experiment results on datasets of TRECVID 2005 and Caltech-256 show that our method is superior to the state-of-the-art multiple kernel learning methods.