Improving face clustering using social context

In this paper we describe an algorithm to improve the performance of face clustering using the social relationship of people. One common challenge in face clustering techniques is that very often the faces of the same person are clustered into different face clusters, due to the imperfection of the face features. The remedy to this problem, the user needs to scan all the clusters and manually merge the face clusters of the same person to the same cluster. We propose to use the social context information inherent among the people in a collection to build a social network and combine this knowledge with face similarity measure to generate a small number of ranked face clusters as the candidate for a cluster to be merged to. Thus, a user can gain the benefit of often avoiding browsing the face clusters back and forth to find the right cluster to merge. Experimental results show that the proposed approach can improve the recall of face clustering because more correct faces are merged into their significant cluster while still maintaining a high precision.