We propose OmniGaze, a semi-supervised learning framework, which utilizes large-scale unlabeled data and reward-driven pseudo-labeling strategy to effectively generalize gaze estimation in the wild.