Gaze in the Dark: Gaze Estimation in a Low-Light Environment with Generative Adversarial Networks

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Accurate user intent prediction is vital for natural interaction in smart environments like digital museums.
• Gaze estimation is a key technology for understanding user intent, outperforming other methods like face orientation and body tracking.
• Existing gaze estimation methods fail in low-light conditions, limiting their practical application.

Purpose of the Study:

• To develop a robust gaze estimation technique effective under various low-light conditions.
• To enhance the practicality of gaze estimation for real-world smart interaction scenarios.

Main Methods:

• A novel deep learning approach combining generative adversarial networks (GANs) and convolutional neural networks (CNNs).
• GANs are used to restore and enhance eye images captured in low-light environments.
• The enhanced images are then processed by a CNN for accurate gaze direction estimation.

Main Results:

• The proposed method significantly improves gaze estimation performance in low and dark light conditions.
• An average performance improvement of 4.53%-8.9% was observed on a modified MPIIGaze dataset.
• Demonstrated the effectiveness of GANs in preprocessing low-light eye images for improved gaze analysis.

Conclusions:

• The developed deep learning approach offers a practical solution for gaze estimation in challenging low-light environments.
• This research represents a significant advancement towards more natural and intuitive human-computer interaction in smart settings.
• Further research is warranted to explore the full potential of this technique in diverse interactive applications.