Towards Accurate, Computationally Efficient Appearance-Based Gaze Estimation via Eyes-Only Pipeline with Additional Regularization
Jan Glinko · 2025
Accurate and computationally efficient gaze estimation is crucial for applications in robotics, autonomous systems, and human-computer interaction. Existing deep-learning models for appearance-based gaze estimation often require high computational resources and large input images, making them impractical for resource-constrained environments. This work introduces an optimized eyes-only pipeline, reducing input image size while maintaining or improving accuracy. Our approach enhances generalization without increasing inference complexity by integrating Regularization Branches and Regression Focal Loss. On the GazeCapture dataset, we achieve angular errors of$3.43^{\circ}, {3.23^{\circ}}$, and 3.15° for input images of size$112 \times 112,224 \times 224$, and$448 \times 448$, surpassing the full-face baseline by$0.48^{\circ}, 0.26^{\circ}$, and 0.23°, respectively. Moreover, our method enables a 16 times reduction in input size while preserving the accuracy of the full face pipeline, demonstrating its effectiveness for gaze estimation on resource-constrained edge devices.