Adaptive Gaze Estimation With Extended Memorization
Jan Glinko · IEEE Access · 2025
Appearance-based gaze estimation still lacks generalizability due to anatomical and appearance differences at the human level. Their performance deteriorates significantly during cross-person evaluation and even more during cross-dataset evaluation. These drawbacks can be mitigated by employing subject-specific calibration data to personalize gaze estimation networks. Here, we present Adaptive Gaze Estimation with Extended Memorization (GEM), an adaptive gaze estimation framework. The core of GEM is a few-shot learning system and novel Extended Memorization. As a successful few-shot learner, GEM adapts to any previously unseen subject without overfitting, while Extended Memorization increases the efficacy and stability of the adaptation process. Crucially, GEM avoids the performance plateau seen in other methods; its accuracy improves consistently as calibration samples increase from 1 to 256. This allows it to compete with and quickly outperform prior deep-learning-only approaches with very few samples, and ultimately achieve state-of-the-art performance by continuing to improve as more calibration data is provided. GEM achieves the state-of-the-art performance of 2.32° on the MPIIFaceGaze dataset and 1.68° on the GazeCapture dataset during the cross-person evaluation.