Hierarchical Tree Representation Based Face Clustering for Video Retrieval

We present a video as a set of people, each person is a sequence of faces clustered by proposed hierarchical tree representation with the purpose of finding all the occurrences of a person in the video without any help of textual information. In the proposed method, faces in a video are detected and tracked to be face-tracks at first, and each face-track is associated with one person. By leveraging temporal constrains, face-tracks that depict the same person in a video are connected. Then we build undirected graphs for a video, and extend discriminative histogram intersection metric learning to generate semantic distances for cutting undirected graphs to be face clusters without predefining the number of clusters. When searching for videos containing the person of query, it is effective to compare faces of query video with sets of people summarized from videos in the dataset. Experimental results show that the proposed face clustering can improve the mean Average Precision of video retrieval and decrease the query time compared to several state-of-the-art approaches.