A Novel Unsupervised 2-Stage k-NN Re-Ranking Algorithm for Image Retrieval

In many image retrieval systems, re-ranking is an important final step to improve the retrieval accuracy given an initial ranking list. K-Nearest Neighbors (k-NN) re-ranking algorithms are the class of algorithms that re-rank an initial ranked list by comparing the similarity between a query image's k-NN and the k-NN of candidate database images, e.g. the initially high ranked images. In this paper, we present a novel 2-stage k-NN re-ranking algorithm. In stage one, we generate an expanded list of candidate database images for re-ranking so that some lower ranked ground truth images will be included for the next stage. In stage two, we re-rank the list of candidate images using a confidence score which is calculated based on both the ranking consistency and reciprocal k-NN properties. Our experimental results on two popular benchmark datasets along with a large-scale 1 million distraction dataset show improved performance over existing k-NN re-ranking methods.