A video face clustering approach based on sparse subspace representation

Abstract

With the changes of illumination, action and background, face clustering is a challenging task that demands accuracy and robustness. In order to improve the face clustering performance in videos, we propose a method which considers the available prior knowledge, multi-view and constrained information. First, multiple features of images are extracted, and sparse subspace clustering algorithm is used to achieve the coefficient matrix. Then, the constrained track matrix and KNN are used to reconstruct the coefficient matrix. Finally, the clustering result is obtained by co-training spectral clustering. The experiment results on two real-world video datasets demonstrate the effectiveness of the approach.