ArbGaze: Gaze Estimation from Arbitrary-Sized Low-Resolution Images.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Gaze estimation aims to determine a gaze vector from facial images.
• Current methods often struggle with varying image resolutions common in real-world scenarios.
• Resolution variations degrade the performance of existing gaze estimation models.

Purpose of the Study:

• To propose a robust gaze estimation method for arbitrary-sized, low-resolution images.
• To address the performance degradation caused by resolution variations in in-the-wild images.
• To improve the generalizability of gaze estimation techniques.

Main Methods:

• A novel gaze estimation approach combining knowledge distillation and feature adaptation.
• Knowledge distillation is used to generate feature maps comparable to high-resolution images.
• Feature adaptation enables processing of diverse image resolutions by integrating low-resolution images with scale information.

Main Results:

• The proposed method significantly improves gaze estimation performance in ablation studies.
• Combining knowledge distillation and feature adaptation yields substantial performance gains.
• The approach demonstrates effectiveness across various backbone architectures, indicating strong generalizability.

Conclusions:

• The developed method effectively handles arbitrary-sized low-resolution images for gaze estimation.
• The integration of knowledge distillation and feature adaptation offers a promising solution for real-world gaze analysis.
• This technique enhances the robustness and applicability of gaze estimation models in diverse environments.