Topic Models for Image Localization

We present a new scheme for partitioning geo-tagged reference image database in an effort to speed up (query) image localization while maintaining acceptable localization accuracy. Our method learns a topic model over the reference database, which in turn is used to divide the reference database into scene groups. Each scene group consists of “visually similar” images as determined by the topic model. Next raw Scale-Invariant Feature Transform (SIFT) features are collected from every image in a scene group a Fast Library for Approximate Nearest Neightbours (FLANN) index is constructed. Given a query image, first its scene group is determined using the topic model and then its SIFT features are matched against the corresponding FLANN index. The query image is localized using the location information from the visually similar images in the reference database. We evaluate our approach on Google Map Street View dataset and demonstrate that our method outperforms a competing technique.