Fast and Memory Saving Instance Search with Approximate Reverse Nearest Neighbor Search Using Reverse Lookup

Abstract

Recently more and more videos have been shared through the websites such as youtube.com. In order to utilize them efficiently, instance search (INS) techniques which find a specific person, object and place from a video database without metadata has been desired. It is known that the BM25 scoring method is a powerful tool for the INS task. It is, however, also known that it requires a time consuming process. It has been pointed out that the time consuming process is equivalent to the bichromatic reverse nearest neighbor (BRNN) search problem and a method to approximate it has been proposed. However, the algorithm needs a huge reference table which causes large memory usage and computational complexity. In this paper, we propose a more efficient way to search BRNNs using a reverse lookup structure. An experimental result using the database of TRECVID 2012 INS task showed that the proposed method was 2.8 times faster with 10% less memory usage than the conventional method.