“Blinks in the Dark”: Blink Estimation with Domain Adversarial Training (BEAT) Network

Blink detection plays an important role in many human-computer interaction applications for consumers. Unfortunately, deep neural network-based blink methods are not only susceptible to poor lighting conditions, but also the learning model is prone bias due imbalance dataset distribution. To solve these problems, we propose Estimation with Domain Adversarial Training (BEAT) network, which robustly detects blinks unseen out-of-sample images captured even under conditions by extracting domain-invariant features. BEAT network inspired domain-adversarial (DANN) improved several improvements including a lambda scheduler stabilize adversarial training and gradient decay layer prevent discriminative loss from overwhelming classification loss. As result, achieves faster more accurate performances than other domain generalization target domains. In particular, BEAT’s feature extractor state-of-the-art performance terms of AUPR on popular benchmark datasets. Also, suggest practical optimal threshold based our insights gained experiments consumer applications.