Gaze-GZ: Generalized Gaze Estimation with Multi-scale Gaze Zone Prediction

Gaze estimation models often experience significant performance degradation on cross-domain tests. Existing methods enforce the model to concentrate on isolating gaze-pertinent features by filtering out irrelevant ones. This paper proposes an advanced generalized framework for gaze estimation, which employs multi-scale gaze zone prediction as an auxiliary task to improve generalization capabilities. Specifically, each facial image is assigned a discrete zone sub-label, alongside the continuous gaze direction label. In addition, we introduce the triplet loss module and the feature consistency branch to ensure that the extracted features within each zone maintain ordered embeddings and robustness to the environmental variations, respectively. The results from comprehensive experiments reveal that the proposed method outperforms all the state-of-the-art generalized gaze estimation methods. The code will be available at github.