Scholarlink

<p>Convolutional neural networks demonstrate strong capability in local spatio-temporal modeling, but they have limitations in capturing global dependency. The transformer architecture has a strong ability to model global dependency but cannot effectively capture local features. For fine-grained action recognition,the combination of the two is beneficial to extract richer fine-grained features. To this end, we introduce a dual cross transformer framework. The multi-scale features from different levels of the convolutional network are extracted and divided into low-level features and high-level features, which are then fed into the dual cross transformer module. The dual cross transformer enables the information interaction of low-level features and high-level features, fuses multi-scale features and models global dependency. Our model achieves competitive performance across various challenging datasets, including FineGym and Diving48,highlighting its strong capability in modeling complex temporal data.</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

Dual cross transformer based on multi-scale fusion for fine-grained action recognition

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Jirui Di,
  • Zhengping Hu,
  • Hehao Zhang,
  • Qiming Zhang,
  • Zhe Sun

摘要

Convolutional neural networks demonstrate strong capability in local spatio-temporal modeling, but they have limitations in capturing global dependency. The transformer architecture has a strong ability to model global dependency but cannot effectively capture local features. For fine-grained action recognition,the combination of the two is beneficial to extract richer fine-grained features. To this end, we introduce a dual cross transformer framework. The multi-scale features from different levels of the convolutional network are extracted and divided into low-level features and high-level features, which are then fed into the dual cross transformer module. The dual cross transformer enables the information interaction of low-level features and high-level features, fuses multi-scale features and models global dependency. Our model achieves competitive performance across various challenging datasets, including FineGym and Diving48,highlighting its strong capability in modeling complex temporal data.