Person-Specific Gaze Estimation from Low-Quality Webcam Images.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Gaze estimation is crucial for applications in healthcare, virtual reality, and human-computer interaction.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has shown significant success in various computer vision tasks.
• Recent advancements have focused on applying deep learning to gaze estimation.

Purpose of the Study:

• To develop and evaluate a person-specific gaze estimation model using CNNs.
• To assess the performance of person-specific models against generalized models.
• To demonstrate the feasibility of using low-quality webcam images for accurate gaze estimation.

Main Methods:

• A person-specific gaze estimation model was implemented using a CNN.
• A dataset of face and eye images was collected using a standard desktop webcam.
• Various CNN hyperparameters, including learning and dropout rates, were tested and optimized.

Main Results:

• Person-specific models demonstrated superior performance compared to generalized models.
• The best Mean Absolute Error (MAE) achieved was 30.09 for the whole face (1.14 degrees).
• Specific MAE results were: left eye 38.20 (1.45 degrees), right eye 36.01 (1.37 degrees), and both eyes 51.18 (1.98 degrees).

Conclusions:

• Person-specific gaze estimation models, when properly hyperparameter-tuned, yield better results than universal models.
• The proposed method is effective even with low-quality images from standard webcams.
• This approach offers a practical and hardware-independent solution for gaze estimation.