Proxy Tuning for Gaze Tracking: A Comparative Study

This study explores proxy tuning as an efficient alternative to traditional fine-tuning for gaze tracking in computer vision. Proxy tuning leverages small expert and anti-expert models to adjust output logits of large pretrained models without modifying their parameters, reducing computational overhead while maintaining competitive accuracy. Using the Columbia Gaze Dataset, I compare proxy tuning against fine-tuning and ablation models, evaluating performance in horizontal and vertical gaze estimation. Results demonstrate proxy tuning achieves an 83% improvement in Mean Squared Error (MSE) for horizontal gaze tasks while significantly lowering resource requirements. These findings highlight proxy tuning's potential for scalable, real-world gaze tracking applications. Project page: https://github.com/GChenCeph/eye_tracking