A Domain Generalization Approach Based on Cost-sensitive Learning for Gaze Estimation

Appearance-based gaze estimation methods regress gaze directions from face images. Deep learning has become the dominant approach to appearance-based gaze estimation and has achieved promising performance within individual datasets. However, gaze estimation methods based on deep learning still perform poorly in cross-domain scenarios. In this paper, we propose a cost-sensitive learning approach for gaze estimation domain generalization. As is observed during cross-domain testing, estimated gazes with large deviations tend to cluster around regions with dense labels in the source domain. Addressing this issue has the potential to improve the generalization ability of the gaze estimation model. To achieve this, we assign weights to the losses generated by each training sample. These weights are determined by two factors: the distribution of the training samples and the rarity of the gaze direction. Experiment show that our method, without any additional network parameters, achieves state-of-the-art performance in gaze estimation domain generalization tasks and reduces overall deviation in cross-domain testing. It is even competitive compared to unsupervised domain adaptation methods for gaze estimation.