TY - GEN
T1 - Towards reading trackers in the wild
T2 - 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and ACM International Symposium on Wearable Computers, UbiComp/ISWC 2017
AU - Ishimaru, Shoya
AU - Hoshika, Kensuke
AU - Kise, Koichi
AU - Dengel, Andreas
AU - Kunze, Kai
N1 - Funding Information:
This work is supported by JST CREST and JSPS KAKENHI (Grant Numbers: JPMJCR16E1, 17K12728).
PY - 2017/9/11
Y1 - 2017/9/11
N2 - 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. Copyright held by the owner/author(s).
AB - 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. Copyright held by the owner/author(s).
KW - Convolutional neural network
KW - Electrooculography
KW - Eye movement
KW - Quantified self
KW - Reading
KW - Recurrent neural network
UR - http://www.scopus.com/inward/record.url?scp=85030867676&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85030867676&partnerID=8YFLogxK
U2 - 10.1145/3123024.3129271
DO - 10.1145/3123024.3129271
M3 - Conference contribution
AN - SCOPUS:85030867676
T3 - UbiComp/ISWC 2017 - Adjunct Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM International Symposium on Wearable Computers
SP - 704
EP - 711
BT - UbiComp/ISWC 2017 - Adjunct Proceedings of the 2017 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2017 ACM International Symposium on Wearable Computers
PB - Association for Computing Machinery, Inc
Y2 - 11 September 2017 through 15 September 2017
ER -