Multiview face retrieval in surveillance video by active training sample collection

For multiview face retrieval of certain person in surveillance video, a key challenge is the lack of training samples. Generally, the law enforcement agencies usually have only one frontal view face image of the target person, however, the faces of the target person in the surveillance video could be in different orientation, and it is impossible for a classifier trained on only frontal view face to retrieve the faces under other orientation. This paper proposes an active training sample collection method for multiview face retrieval in surveillance video. First, the front view face image is used to train a classifier to retrieve the target person's front view face in videos. As the video is continuous, we can track the face and obtain side view faces of the target person. Then these selected side view faces are combined with the frontal view face to form a new training data set. The classifier is updated based on the new training data set, and can retrieve multiview faces of the target people. The experimental results prove the effectiveness of the proposed method.