Scholarlink

Precise Event Spotting (PES) is essential in fast-paced sports such as tennis, where fine-grained events occur within very short temporal windows. Accurate frame-level localization is challenging due to motion blur, subtle action differences, and limited annotated data. We study two complementary distillation strategies for few-shot PES: 1) Adaptive Weight Distillation (AWD), a prediction-level method that adaptively weights teacher supervision on unlabeled data, and 2) Annealed Multimodal Distillation for Few-Shot Event Detection (AMD-FED), a representation-level framework that transfers robust skeleton knowledge into visual modalities through annealed pseudo-labeling. Both methods leverage multimodal distillation to improve generalization under limited supervision. Experiments on \(F^3\) Set-Tennis(sub) under two few-shot settings, 50-clip and 100-clip, show that our methods consistently outperform single-modality baselines and previous PES approaches. In particular, AMD-FED achieves the best overall performance, reaching 80.24% and 83.84% Edit score in the 50-clip and 100-clip settings, respectively, exceeding the strongest single-modality baseline by 9.17 and 7.19% points.

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

From Skeletons to Pixels: Few-Shot Precise Event Spotting via Representation and Prediction Distillation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhong Han Ervin Yeoh,
  • Kan Jiang

摘要

Precise Event Spotting (PES) is essential in fast-paced sports such as tennis, where fine-grained events occur within very short temporal windows. Accurate frame-level localization is challenging due to motion blur, subtle action differences, and limited annotated data. We study two complementary distillation strategies for few-shot PES: 1) Adaptive Weight Distillation (AWD), a prediction-level method that adaptively weights teacher supervision on unlabeled data, and 2) Annealed Multimodal Distillation for Few-Shot Event Detection (AMD-FED), a representation-level framework that transfers robust skeleton knowledge into visual modalities through annealed pseudo-labeling. Both methods leverage multimodal distillation to improve generalization under limited supervision. Experiments on \(F^3\) Set-Tennis(sub) under two few-shot settings, 50-clip and 100-clip, show that our methods consistently outperform single-modality baselines and previous PES approaches. In particular, AMD-FED achieves the best overall performance, reaching 80.24% and 83.84% Edit score in the 50-clip and 100-clip settings, respectively, exceeding the strongest single-modality baseline by 9.17 and 7.19% points.