Scholarlink

This study introduces a video captioning model designed for classroom surveillance videos, particularly those recorded in computer labs where student behaviors are often subtle, such as sitting still, focusing, using devices, brief interactions, or moving within the room. The model integrates visual features from VideoMAE and CLIP across three branches—Action Recognition, Object Identification, and Contextual Understanding—which are fused by Q-Former and decoded with DeepSeek-VL2Evaluated on a real-world classroom dataset, the model generates concise behavior descriptions and outperforms traditional baselines in capturing semantically meaningful micro-level behaviors, achieving a BLEU-4 score of 0.410 and a CIDEr score of 0.655.

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

Q-ClassCap: An End-to-End Structured Captioning Model for Classroom Activity Videos

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nguyen Dinh Quy,
  • Tran Manh Tuan,
  • Le Hoang Son

摘要

This study introduces a video captioning model designed for classroom surveillance videos, particularly those recorded in computer labs where student behaviors are often subtle, such as sitting still, focusing, using devices, brief interactions, or moving within the room. The model integrates visual features from VideoMAE and CLIP across three branches—Action Recognition, Object Identification, and Contextual Understanding—which are fused by Q-Former and decoded with DeepSeek-VL2Evaluated on a real-world classroom dataset, the model generates concise behavior descriptions and outperforms traditional baselines in capturing semantically meaningful micro-level behaviors, achieving a BLEU-4 score of 0.410 and a CIDEr score of 0.655.