Towards reading trackers in the wild

Reading in real life occurs in a variety of settings. One may read while commuting to work, waiting in a queue or lying on the sofa relaxing. However, most of current activity recognition work focuses on reading in fully controlled experiments. This paper proposes reading detection algorithms that consider such natural readings. The key idea is to record a large amount of data including natural reading habits in real life (more than 980 hours from 7 participants) with commercial electrooculography (EOG) glasses and to use them for deep learning. Our proposed approaches classified controlled reading vs. not reading with 92.2% accuracy on a user-dependent training. However, the classification accuracy decreases to 73.8% on natural reading vs. not reading. The results indicate that there is a strong gap between controlled reading and natural reading, highlighting the need for more robust reading detection algorithms.