Two methods to efficiently train kernelized support vector machines are introduced. Both of them apply stochastic gradient descent in the primal space. Different from previous fast stochastic kernel machines method [9] which drops old support vectors directly, one of the algorithms exploits the efficient representation of the histogram intersection kernel, the other one approximates the discarded support vectors with existing ones. The experiments are conducted on PASCAL VOC 2007 dataset [4]. The experiments show that our methods use significantly less training time than the batch training method. The algorithms are also tested on a dataset with about 105 training images for each image category. The results show that the classification performance is consistently improved by increasing training data size. Efficiently training kernel machines with giant image datasets is a promising way to do image classification.
展开▼
