Unsupervised Outlier Detection in Appearance-Based Gaze Estimation

Zhaokang Chen, Didan Deng, Jimin Pi, Bertram E. Shi; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 0-0


Appearance-based gaze estimation maps RGB images to estimates of gaze directions. One problem in gaze estimation is that there always exist low-quality samples (outliers) in which the eyes are barely visible. These low-quality samples are mainly caused by blinks, occlusions (e.g. by eye glasses), blur (e.g. due to motion) and failures of the eye landmark detection. Training on these outliers degrades the performance of gaze estimators, since they have no or limited information about gaze directions. It is also risky to give estimates based on these images in real-world applications, as these estimates may be unreliable. To solve this problem, we propose an algorithm that detects outliers without supervision. Based on the input images with only gaze labels, the proposed algorithm learns to predict a gaze estimates and an additional confidence score, which alleviates the impact of outliers during learning. We evaluated this algorithm on the MPIIGaze dataset and on an internal dataset. In cross-subject evaluation, our experimental results show that the proposed algorithm results in a better gaze estimator (8% improvement). The proposed algorithm is also able to reliably detect outliers during testing, with a precision of 0.71 when the recall is 0.63.

Related Material

author = {Chen, Zhaokang and Deng, Didan and Pi, Jimin and Shi, Bertram E.},
title = {Unsupervised Outlier Detection in Appearance-Based Gaze Estimation},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
month = {Oct},
year = {2019}