Large Generative Model Impulsed Lightweight Gaze Estimator via Deformable Approximate Large Kernel Pursuit.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Lightweight gaze estimation is crucial for mobile AR/VR applications.
• Current deep learning models face challenges with computational cost and generalization.
• Existing methods struggle with diverse eye appearances and subtle pupil movements.

Purpose of the Study:

• To develop an efficient and accurate lightweight gaze estimation method.
• To address the limitations of heavy computational architectures and poor generalization in current models.
• To enable robust gaze tracking on resource-constrained mobile platforms.

Main Methods:

• Proposed a novel lightweight network with a deformable approximate large kernel.
• Extended the receptive field to handle complex eye movements and appearance variations.
• Integrated gaze estimation training with a control information extraction module.
• Utilized a large generative model (Stable Diffusion V1.5) for gaze-specific image generation.
• Distilled generalization capabilities from the generative model into the lightweight gaze model.

Main Results:

• The proposed model demonstrates superior accuracy compared to state-of-the-art methods.
• Achieved high performance with significantly reduced model complexity.
• Effectively handles diverse eye textures and subtle pupil movements.
• Validated the effectiveness of the deformable kernel and generative model integration.

Conclusions:

• The novel lightweight network offers an efficient and accurate solution for gaze estimation.
• The proposed training scheme successfully distills generalization from large generative models.
• This method is suitable for deployment on mobile devices and AR/VR systems.
• Represents a significant advancement in real-time, on-device gaze tracking technology.