Fusing Appearance Features and Correlation Features for Face Video Retrieval

摘要

Face video retrieval has drawn considerable research attention recently. Most prior research mainly focused on either appearance features or correlation features, which could degrade retrieval performance. In this paper, we fuse appearance features and correlation features to exploit rich information of face videos for face video retrieval via a deep convolutional neural network. The network extracts appearance feature and correlation feature from a frame and the covariance matrix of a face video, respectively, and fuses them to obtain a comprehensive video representation. The fused feature is projected to a low-dimensional Hamming space via hash functions for the retrieval task. The network integrates feature extractions, feature fusion, and hash learning into a unified optimization framework to guarantee optimal compatibility of appearance features and correlation features. Experiments on two challenging TV-Series datasets demonstrate the effectiveness of the proposed method.