Efficient Content-based Image Retrieval for Position Estimation on GPU

We propose an efficient content-based image retrieval (CBIR) method for position estimation of mobile devices. The idea is to use videos of first-person vision associated with geographical position information as the database. When a user sends a current subjective image, the system estimates the position using CBIR. Since features extracted from images are in general high-dimensional vectors, thousands of vectors are extracted even from a single image, resulting in high processing cost. To tackle this problem, we have proposed a method in which features are compressed using LSH, and GPU is used for accelerating CBIR. Nevertheless, it suffered from performance degradation due to write conflicts among different threads. This paper presents an improved method which avoids write conflicts and modifies LSH algorithm to improve accuracy. Also, we demonstrate the efficiency and accuracy of the proposed scheme through experiments using a video dataset.