Web Image Gathering with a Spatial Pyramid Kernel

In this paper, we propose a Web image gathering system em- ploying state-of-the-art object recognition techniques, which are bag-of-visual-words representation and a spatial pyramid kernel. In these several years, research on object recognition is progressing greatly. Most of work on object recognition assumes complete training data is available, while complete training data is not available in general in case of gathering Web images with no human intervention. The objective of this paper is to examine if a state-of-the-art object recogni- tion technique is also effective for the Web image gathering task where training images always include some noise. By the experiments, we show the state-of-the-art object recognition method is also very effective for Web image gathering and the results outperform ones by existing methods greatly.