Uncertainty Modeling for Gaze Estimation.

Area of Science:

• Computer Vision
• Medical Research
• Machine Learning

Background:

• Gaze estimation is crucial for computer vision and medical applications.
• Existing methods often neglect input and annotation uncertainties, leading to deterministic predictions.
• This limitation hinders the practical applicability of current gaze estimation models.

Purpose of the Study:

• To develop a probabilistic framework for gaze estimation that accounts for input and annotation uncertainties.
• To introduce instance-wise uncertainty estimation for confidence measurement in predictions.
• To enhance the accuracy and reliability of gaze estimation in real-world scenarios.

Main Methods:

• Probabilistic embeddings are used to model input uncertainty, representing images as Gaussian distributions.
• A novel label distribution learning method, probabilistic annotations, models annotation uncertainty using Gaussian distributions.
• An Embedding Distribution Smoothing (EDS) module and hard example mining are developed to improve distribution consistency.

Main Results:

• The proposed probabilistic framework significantly improves gaze estimation accuracy.
• The method outperforms baseline and state-of-the-art approaches on benchmark datasets (GazeCapture, MPIIFaceGaze) and mobile data.
• Instance-wise uncertainty estimation provides critical confidence measures for predictions.

Conclusions:

• The probabilistic framework effectively models and addresses uncertainties in gaze estimation.
• This approach offers more reliable and interpretable gaze prediction results.
• The findings have significant implications for advancing computer vision and medical research in eye-tracking.