Gaze Estimation Based on Convolutional Structure and Sliding Window-Based Attention Mechanism.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Human gaze direction is crucial for understanding attention and cognitive states.
• Convolutional Neural Networks (CNNs) show limitations in global modeling for gaze estimation.
• Transformer models offer advancements but can degrade local spatial information.

Purpose of the Study:

• To introduce Swin Transformer architectures for enhanced gaze estimation.
• To address limitations of existing CNN and Transformer models in capturing multiscale features and local spatial details.

Main Methods:

• Developed two Swin Transformer-based gaze estimation models: SwinT-GE (pure Swin Transformer) and Res-Swin-GE (hybrid CNN-Swin Transformer).
• Res-Swin-GE integrates convolutional structures to replace the slicing-and-mapping mechanism in SwinT-GE.
• Evaluated models on the MpiiFaceGaze and Eyediap datasets.

Main Results:

• Res-Swin-GE significantly outperformed the pure SwinT-GE model.
• The hybrid Res-Swin-GE model demonstrated strong competitiveness on the MpiiFaceGaze dataset.
• Achieved a 7.5% performance improvement over state-of-the-art methods on the Eyediap dataset.

Conclusions:

• Swin Transformer architectures, particularly the hybrid Res-Swin-GE, offer a promising direction for advancing gaze estimation.
• The hybrid approach effectively balances global and local feature learning for superior performance.
• This work contributes to more accurate analysis of human behavior and cognitive states through gaze tracking.