Scholarlink

<p>Understanding students’ behavior is vital for analyzing their engagement in the classroom. Benefiting from AI technology, researchers are striving to model the behaviors by students in-class actions and present some public benchmarks. However, these datasets usually classify students’ actions into exclusive categories, ignoring complex interactions and finer details of behaviors. Other researchers attempt to understand students’ subtle behaviors through their attention, but facing difficulties in accurately annotating 3D gaze orientations from 2D images. To address above issues, we propose a novel fine-grained class students’ behavior understanding dataset, called FineEdu. This dataset takes into account both actions and attention, focusing on three crucial cues: <i>“where are they”</i>, <i>“what are they doing"</i> and <i>“what are they looking at”</i>. Compared with previous action datasets, we divide action categories into two super-categories: <i>posture</i> and <i>action</i> categories to keep the atomicity of the dataset. Additionally, we utilize <i>gaze objects</i> as annotations of attention instead of 3D orientations, making more semantic meaning and easier to accurately annotate. It is the initial public dataset to consider actions and gaze objects jointly for students’ behavior understanding. We have collected 103 videos and annotated over 5,000 images from 82 classrooms. To show its diversity, reality, and challenges, we systematically investigate the characteristics of FineEdu and implement representative methods on this dataset, obtaining a number of observations. This dataset is released on <a href="https://github.com/ZZZZZZZZJ/FineEdu-Dataset">https://github.com/ZZZZZZZZJ/FineEdu-Dataset</a>. We hope this dataset could advance research towards class students’ behavior understanding in education applications.</p>

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

FineEdu: a fine-grained class students behavior understanding dataset with jointly action and attention annotations

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Zhijun Zhang,
  • Ziyue Feng,
  • Zhiying Yan,
  • Xu Zou,
  • Sheng Zhong,
  • Zheng Zhang

摘要

Understanding students’ behavior is vital for analyzing their engagement in the classroom. Benefiting from AI technology, researchers are striving to model the behaviors by students in-class actions and present some public benchmarks. However, these datasets usually classify students’ actions into exclusive categories, ignoring complex interactions and finer details of behaviors. Other researchers attempt to understand students’ subtle behaviors through their attention, but facing difficulties in accurately annotating 3D gaze orientations from 2D images. To address above issues, we propose a novel fine-grained class students’ behavior understanding dataset, called FineEdu. This dataset takes into account both actions and attention, focusing on three crucial cues: “where are they”, “what are they doing" and “what are they looking at”. Compared with previous action datasets, we divide action categories into two super-categories: posture and action categories to keep the atomicity of the dataset. Additionally, we utilize gaze objects as annotations of attention instead of 3D orientations, making more semantic meaning and easier to accurately annotate. It is the initial public dataset to consider actions and gaze objects jointly for students’ behavior understanding. We have collected 103 videos and annotated over 5,000 images from 82 classrooms. To show its diversity, reality, and challenges, we systematically investigate the characteristics of FineEdu and implement representative methods on this dataset, obtaining a number of observations. This dataset is released on https://github.com/ZZZZZZZZJ/FineEdu-Dataset. We hope this dataset could advance research towards class students’ behavior understanding in education applications.