Large scale specific object recognition by using GIFTS image feature

Abstract

We propose GIFTS (Goods Image Features for Tree Search) which uses image local features for large-scale object recognition. Each GIFTS is a kind of keypoint feature. The feature vector consists of intensity deltas for 128 selected pixel pairs around the keypoint. By generating a KD-Tree from the GIFTS feature vectors of the training images and using the KD-Tree to search for nearest neighbor feature vectors of a query image, query times are on the order of log N for specific object recognition. We used the proposed method for book cover queries with 100,000 training images, had recognition accuracy over 99% with query times within one second.