Efficient re-ranking in vocabulary tree based image retrieval

Image retrieval using a large vocabulary tree of local invariant features can efficiently handle databases with millions of images. However, a costly re-ranking step is generally required to re-order the top candidate images to enforce spatial consistency among local features. In this paper, we propose an efficient re-ranking approach which takes advantage of the vocabulary tree quantization to conduct fast feature matching. The proposed re-ranking algorithm involves no operations in the high-dimensional feature space and does not assume a global transform between a pair of images, thus, it not only dramatically reduces the computational complexity but also improves the retrieval precision, which is validated using 1.26 million images in the public ImageNet dataset and the San Francisco Landmark dataset including 1.7 million images.