Self-Adaptive Appearance-Based Eye-Tracking with Online Transfer Learning
Bruno Klein Salvalaio, Gabriel de Oliveira Ramos · 2019
Eye-tracking plays a role in human-computer interactions and has proven useful in a wide variety of domains. We consider appearance-based eye-tracking, where one tracks eye movements based solely on conventional images (rather than on sophisticated additional hardware). Recent advances made in Deep Learning and, in particular, convolutional neural networks have allowed appearance-based eye-tracking to achieve better results than ever. However, current literature still lacks methods that generalize to different combinations of user, environment and device. In this work, we introduce Online Deep Appearance-Based Eye-Tracking (ODABE), which overcomes such a limitation by considering online transfer learning, thus enabling eye-tracking models to self-adapt to different context very rapidly. Our results show that ODABE improves upon previous research when context changes, decreasing the prediction error by 50.95% on average, on tested cases.