Scholarlink

With the development of immersive and interactive virtual environments, accurately estimating the users’ gazes could enhance the evaluation of visual design or gaze-driven interaction, so the precise estimation of gaze fixation has become increasingly crucial. This paper proposes a near-eye gaze estimation method inspired by the Res-Net network and incorporates eye appearance features and bottle-net attention module (BAM) to calculate 2D gaze fixation coordinates. We build a homemade cardboard box-based VR headset by using a mobile phone along with two infrared cameras. Finally, we conducted a user study, and the results show that the proposed method reduces gaze point coordinates’ error to 15.3 pixels and visual angle error to 3.54˚. The results outperform existing methods and will leverage gaze point interaction in virtual reality applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Near-Eye Gaze Estimation in Virtual Reality Based on Deep Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zian Sun,
  • Yang Liu,
  • Shiwei Cheng

摘要

With the development of immersive and interactive virtual environments, accurately estimating the users’ gazes could enhance the evaluation of visual design or gaze-driven interaction, so the precise estimation of gaze fixation has become increasingly crucial. This paper proposes a near-eye gaze estimation method inspired by the Res-Net network and incorporates eye appearance features and bottle-net attention module (BAM) to calculate 2D gaze fixation coordinates. We build a homemade cardboard box-based VR headset by using a mobile phone along with two infrared cameras. Finally, we conducted a user study, and the results show that the proposed method reduces gaze point coordinates’ error to 15.3 pixels and visual angle error to 3.54˚. The results outperform existing methods and will leverage gaze point interaction in virtual reality applications.