Neighborhood-Exact Nearest Neighbor Search for face retrieval

Abstract

Features extracted by deep convolutional neural network perform well for face recognition. The crux of large-scale face retrieval using such features stems from the trade-off between speed and accuracy: search efficiently while preserving acceptable accuracy. Traditional methods perform Nearest Neighbor Search (NNS) over the entire gallery set based on feature matching in the continuous space, which is able to achieve exact (optimized) retrieval results for given features. However, it is not feasible for large-scale face retrieval due to unaffordable time consumption. A typical way of speeding up the face retrieval is to conduct Approximate Nearest Neighbor Search (ANNS) at the expense of accuracy without guarantee of search quality. To circumvent this dilemma, we define a new problem w.r.t. large-scale face retrieval, which is referred to as Neighborhood-Exact Nearest Neighbor Search (NENNS). NENNS demands efficient search while guaranteeing the search exactness within a specified neighborhood around the query. Further, we propose a heuristic method for face retrieval, which is able to perform NENNS efficiently. Specifically, it first discretizes the continuous features of samples into binary codes by our designed Angular Binary-Encoding Mechanism. Then it performs fast yet precise exclusion of bad candidates in the binary feature space based on our defined metric (binarized Cosine similarity) to reduce search space significantly. Finally the accurate searching results conforming to NENNS are achieved by conducting precise search in continuous feature space. We perform theoretical proof and extensive experiments to validate the correctness and effectiveness of the proposed method.