Specialized Gaze Estimation for Children by Convolutional Neural Network and Domain Adaptation.

Wen Cui,Jinshi Cui,Hongbin Zha
DOI: https://doi.org/10.1109/icip.2017.8296894
2017-01-01
Abstract:Children's social gaze behavior modeling and evaluation has obtained increasing attentions in various research areas. In psychology research, eye gaze behavior is very important to developmental disorders diagnosis and assessment. In robotics area, gaze interaction between children and robots also draws more and more attention. However, there exists no specific gaze estimator for children in social interaction context. Current approaches usually use models trained with adults' data to estimate children's gaze. Since gaze behaviors and eye appearances of children are different from those of adults, the current approaches, especially those with free calibration assumptions which are utilized in usual human-robot interaction systems, will result in big errors. Note that children data is difficult to collect and label, so directly learning from children data is hard to achieve. We propose a new system to solve this problem, which combines a CNN feature extractor trained from adult data and a domain adaptation unit using geodesic flow kernel to adapt the source domain (adults) classifier to the target domain (children). Our system performs well in children's gaze estimation.
What problem does this paper attempt to address?