Scholarlink

<p>Video action recognition is crucial for video understanding. Traditional methods relying on complete frame features face computational inefficiency and redundancy due to temporal redundancy between frames. Moreover, motion information is vital for accurate recognition but often inadequately captured. To address these challenges, we propose a novel framework with Convolutional Tube Masking (CTM) and Bidirectional Motion Feature Module (BiMFM). CTM uniformly masks features across video frames at a fixed ratio, maintaining temporal continuity while improving computational efficiency. BiMFM models motion from two temporal directions, enabling more accurate and comprehensive motion representation. Furthermore, CTM enhances efficiency by suppressing redundant features, while BiMFM improves recognition performance by reinforcing motion awareness. Finally, we design a reconstruction strategy tailored for convolutional neural networks (CNNs) to maintain the integrity of temporal information. We evaluate our method on three large-scale benchmark datasets: Kinetics-400, Something-Something V1, and V2. Experimental results show that our method achieves satisfactory performance on all datasets.</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

BiCTM: lightweight video recognition via convolutional tube masking and bidirectional motion features

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Linxi Li,
  • Zhengyan Li,
  • Mingwei Tang,
  • Jie Hu,
  • Yanxi Zheng,
  • Qingchi Gui

摘要

Video action recognition is crucial for video understanding. Traditional methods relying on complete frame features face computational inefficiency and redundancy due to temporal redundancy between frames. Moreover, motion information is vital for accurate recognition but often inadequately captured. To address these challenges, we propose a novel framework with Convolutional Tube Masking (CTM) and Bidirectional Motion Feature Module (BiMFM). CTM uniformly masks features across video frames at a fixed ratio, maintaining temporal continuity while improving computational efficiency. BiMFM models motion from two temporal directions, enabling more accurate and comprehensive motion representation. Furthermore, CTM enhances efficiency by suppressing redundant features, while BiMFM improves recognition performance by reinforcing motion awareness. Finally, we design a reconstruction strategy tailored for convolutional neural networks (CNNs) to maintain the integrity of temporal information. We evaluate our method on three large-scale benchmark datasets: Kinetics-400, Something-Something V1, and V2. Experimental results show that our method achieves satisfactory performance on all datasets.