Implementation of Large-Scale Object Recognition System

In this paper we build a simple large-scale object recognition system which consists of several publicly available software. With 1 million distracter image database, we measure precision and search time to show the performance. To support searching within a reasonable time, we need an index structure, e.g. vocabulary tree. We quantize 128 dimensional SIFT feature to a single positive integer value using vocabulary tree. Using simple result refinement step, experimental results show that retrieval accuracy of near 90% precision within less than 3 seconds search time with 1 million image database is achieved.