Scholarlink

<p>Action Quality Assessment (AQA) aims to assess the standard proficiency of individuals in specific actions. Existing AQA datasets focus mostly on fast-paced scenarios, limiting the study of slow, nuanced actions. Therefore, we introduce the first fine-grained Baduanjin (BDJ) AQA dataset to address this gap. Baduanjin, a traditional Chinese fitness exercise known as the Eight-Section Brocade, involves slow and finely nuanced action sequences, challenging AQA models to assess fine motor skills. The dataset comprises 920 videos, each 12 to 19&#xa0;s long, featuring eight distinct actions. Furthermore, we propose a new AQA model, Adaptive Frequency-Aware Network (AFA). This model is the inaugural attempt to introduce the frequency domain into AQA tasks that explore detailed differences between video clips, treating them as interrelated groups rather than independent entities. Adaptive weights are assigned to reflect the varying importance of movements, and score probabilities are predicted through uncertainty-aware distribution learning. AFA achieves outstanding performance on two long-term action datasets, Rhythmic Gymnastics, Fis-V, and our proposed short-term BDJ dataset.</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

Adaptive frequency-aware network for action quality assessment

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Chunting Wang,
  • Xini Ding,
  • Xuan Zhao,
  • Huiliang Shang,
  • Lin Gu,
  • Miao Wang

摘要

Action Quality Assessment (AQA) aims to assess the standard proficiency of individuals in specific actions. Existing AQA datasets focus mostly on fast-paced scenarios, limiting the study of slow, nuanced actions. Therefore, we introduce the first fine-grained Baduanjin (BDJ) AQA dataset to address this gap. Baduanjin, a traditional Chinese fitness exercise known as the Eight-Section Brocade, involves slow and finely nuanced action sequences, challenging AQA models to assess fine motor skills. The dataset comprises 920 videos, each 12 to 19 s long, featuring eight distinct actions. Furthermore, we propose a new AQA model, Adaptive Frequency-Aware Network (AFA). This model is the inaugural attempt to introduce the frequency domain into AQA tasks that explore detailed differences between video clips, treating them as interrelated groups rather than independent entities. Adaptive weights are assigned to reflect the varying importance of movements, and score probabilities are predicted through uncertainty-aware distribution learning. AFA achieves outstanding performance on two long-term action datasets, Rhythmic Gymnastics, Fis-V, and our proposed short-term BDJ dataset.