GSA-Gaze: Generative Self-adversarial Learning for Domain Generalized Driver Gaze Estimation

Estimating driver gaze accurately is critical for the human-machine cooperative driving, but the significant facial appearance diversions caused by background, illumination, personal characteristics, etc. pose a challenge to the generalizability of gaze estimation models. In this paper, we propose the generative self-adversarial learning mechanism for generalized gaze estimation that aims to learn general gaze features while eliminating sample-specific features and preventing cross-domain feature over-fitting. Firstly, to reduce information redundancy, the feature encoder is designed based on pyramid-grouped convolution to extract a sparse feature representation from the facial appearance. Secondly, the gaze regression module supervises the model to learn as many gaze-relevant features as possible. Thirdly, the adversarial image reconstruction task prompts the model to eliminate the domain-specific features. The adversarial learning of the gaze regression and the image reconstruction tasks guides the model to learn only general gaze features across domains, preventing cross-domain feature over-fitting, enhancing the domain generalization capability. The results of cross-domain testing of four active gaze datasets prove the effectiveness of the proposed method. The code is available at https://github.com/HongchengHan/GSA-Gaze