Learning to Attract and Repel: Dual Quality Margin Learning for Face Recognition (DQM-Face) ECCV-26

Introduce DQM-Face, a unified framework for dual quality modeling and dual-margin optimization. Dual quality modeling mechanism: fuses magnitude-based and semantic quality cues for more reliable quality estimation. Dual margin optimization: an adaptive attraction margin (𝑚_1)​ to improve intra-identity compactness, and a scheduled repulsion margin (𝑚_2)​ to increase inter-identity separation. Achieves state-of-the-art (SOTA) face recognition performance across challenging benchmarks, while the learned quality is highly effective for face image quality assessment.