I2DNet - Design and Real-Time Evaluation of Appearance-based gaze estimation system

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Gaze estimation is crucial for human-computer interaction.
• Appearance-based methods, leveraging deep learning, outperform traditional model-based approaches but struggle with cross-dataset generalization.
• Real-time performance and subject independence remain challenges for current state-of-the-art gaze estimation systems.

Purpose of the Study:

• To propose I2DNet, a novel deep learning architecture for improved subject-independent gaze estimation.
• To evaluate I2DNet's performance in a real-time gaze-controlled interface.
• To compare I2DNet against existing systems like Webgazer.js and OpenFace 2.0.

Main Methods:

• Developed I2DNet, a novel deep learning architecture for gaze estimation.
• Achieved state-of-the-art accuracy on MPIIGaze and RT-Gene datasets (4.3 and 8.4 degree MAE).
• Implemented I2DNet as a gaze-controlled interface for a 9-block pointing and selection task, comparing it with Webgazer.js and OpenFace 2.0 in a user study with 16 participants.

Main Results:

• I2DNet achieved state-of-the-art mean angular errors of 4.3° on MPIIGaze and 8.4° on RT-Gene.
• In real-time evaluation, I2DNet significantly reduced selection time compared to Webgazer.js and OpenFace 2.0.
• The proposed system also led to a statistically significant reduction in missed selections.

Conclusions:

• I2DNet demonstrates superior subject-independent gaze estimation accuracy.
• The system offers a robust and efficient gaze-controlled interface for real-time applications.
• I2DNet represents a significant advancement in gaze estimation technology for interactive settings.