SSDL: Self-Supervised Domain Learning for Improved Face Recognition

摘要: Face recognition in unconstrained environments is challenging due to variations illumination, quality of sensing, motion blur and etc. An individual's face appearance can vary drastically under different conditions creating a gap between train (source) varying test (target) data. The domain could cause decreased performance levels direct knowledge transfer from source target. Despite fine-tuning with specific data be an effective solution, collecting annotating for all domains extremely expensive. To this end, we propose self-supervised learning (SSDL) scheme that trains on triplets mined unlabelled A key factor discriminative learning, selecting informative triplets. Building most confident predictions, follow "easy-to-hard" alternate triplet mining self-learning. Comprehensive experiments four benchmarks show SSDL generalizes well domains.