Learn to Cluster Faces with Better Subgraphs

Face clustering can help to improve the performance of face recognition models with unlabeled data. The existing methods generally aggregate the features within subgraphs based on a threshold or learned cutoff positions. This may reduce the recall of subgraphs and hence degrade the clustering performance. This work proposed an efficient neighborhood-aware subgraph adjustment method that can significantly improve the quality of the subgraphs. More specifically, it consists of two components, i.e. feature enhancement embeddings and enclosed subgraph embeddings of node pairs. The embeddings are combined to predict the linkage probabilities for all node pairs to produce new subgraphs that can be further used for aggregation of GCNs or other clustering methods. The numerical results confirm that it outperforms the SOTA solutions in terms of generalization capability.