FAZE: Few-Shot Adaptive Gaze Estimation using Meta-Learning

Intelligent Robotics Seminar: Paper Review & Implementation, Universität Hamburg

This project explores the implementation of Few-Shot Adaptive Gaze Estimation (FAZE), a framework designed to bridge the gap between generic gaze estimation and highly personalized models. Traditional models often struggle with individual anatomical differences, whereas FAZE leverages meta-learning to adapt to new users using as few as 3 to 9 calibration samples.

The core of the system is the Disentangling Transforming Encoder-Decoder (DT-ED) network. This architecture is trained to decouple gaze direction, head pose, and facial appearance from raw images. By utilizing a latent space that represents these features independently, the model can synthesize novel viewpoints through a latent transformation, effectively augmenting the calibration data provided by the user.

A key contribution of this project is the application of MAML (Model-Agnostic Meta-Learning) to the gaze estimation task. By training the network to be highly adaptable, we achieve significant accuracy gains in “personalizing” the model to a specific individual’s eyes and facial features. The results demonstrate that FAZE significantly outperforms static baseline models, providing a robust solution for hands-free interaction and user-intent detection in real-world scenarios.