Find where you are: a new try in place recognition

Abstract

LBS-based applications have become trending on mobile phones. It is useful and necessary to locate users location precisely from a digital image. However, gap exists between the query and the data set in scale, viewpoint, and lighting, or the noise existed in the foreground or background, etc. It is challenging for a location recognition or retrieval system to carry out real-time service. To address this problem, we design a place recognition system and a new building data set with ground truth labels. The algorithm not only significantly improves the efficiency, but also gives satisfied accuracy. The main contributions of our work can be concluded by three points: (1) By adding a fast geometric image matching as a filter procedure before applying Random Sample Consensus (RANSAC), we substantially improve the efficiency of spatial verification and recognition accuracy. (2) We apply a camera orientation algorithm to predict a probable retrieval failure. Salient region detection is applied to remove the confusing features in combination with term frequency---inverse document frequency (tf---idf) recovery. This significantly reduces the influence of noise. (3) We establish a new building data set of Tsinghua University to verify retrieval results. Experiments are conducted on several data sets and all achieve state-of-the-art results.