抄録
Although metadata are useful to obtain better clustering results on image clustering, some images do not have social tags or metadata about photo-taking conditions. In this paper, we propose an image clustering method that is robust for those missing metadata of photo images that appear in search results on the Web. The method has an integrated estimation mechanism for missing social tags or photo-taking conditions from other images in the image search result. An advantage of our method is that our approach does not require another training set that is constructed from other images that are not included in the search result. We demonstrate that the proposed method can effectively cluster images which have some missing metadata by showing the performance of on-demand clustering on a photo sharing site.
