Robust Face Set Recognition with Improved Core Set Selection

Face set recognition is challenging due to the disruptions caused by noise in the sets. Although face recognition has been well explored, there are few methods and specific datasets that focus on the noise problem. Recently proposed core set selection methods have shown their ability to balance sample quality and diversity within sets. Thus, we put forward an innovative approach founded on enhanced core set selection, which is robust against noisy face sets. Specifically, we design an enhanced core set selection by an infomap-based clustering algorithm, which effectively filters out noisy faces. To enhance the feature extraction ability of the model under noisy face sets, this study presents the SimAM attention mechanism into the backbone and propose a memory-friendly feature aggregation module. Our experiment has achieved promising results on several benchmarks (IJB-B, IJB-C, and self-built datasets).