Orthogonal Subspace Decomposition: A New Perspective of Learning Discriminative Features for Face Clustering

Face clustering is an important task, due to its wide applications in practice. Graph-based face clustering methods have recently made a great progress and achieved new state-of-the-art results. Learning discriminative node features is the key to further improve the performance of graph-based face clustering. To this end, most previous methods focus on new loss functions, such as margin-based or center loss. In this paper, we propose subspace learning as a new way to learn discriminative node features, which is implemented by a new orthogonal subspace decomposition (OSD) module. In graph-based face clustering, OSD leads to more discriminative node features, which better reflect the relationship between each pair of faces, thereby boosting the accuracy of face clustering. Extensive experiments show that OSD outperforms state-of-the-art results with a healthy margin.