Local feature reliability measure using multiview synthetic images for mobile visual search

Abstract

In this paper, we propose a new database (DB) construction method for the mobile visual search (MVS) system based on the local feature and bag-of-visual-words framework. In MVS, quantization error is unavoidable and causes performance degradation. Typical approaches for visual search extract features from a single view of reference images, though such features are insufficient to manage the quantization error. In this paper, we generate multiview synthetic images and extract local features. These features are resampled according to our novel reliability measure in order to reduce the DB size. Experiments on the three datasets show that the proposed method successfully constructs a robust DB with same size. The proposed method improved the mean average precision compared with a conventional method without changing the searching procedure.