Scholarlink

<p>Driven by advancements in computer vision and machine learning, automated systems have made remarkable strides in capturing real-time dynamics in table tennis. However, table tennis ball detection still faces numerous challenges. First, the ball’s small size, rapid motion, and variable trajectories result in complex image positions, making accurate detection difficult. Moreover, background interferences such as spectators, billboards, and referees can lead to target confusion, while lighting variations further increase detection uncertainty. These factors collectively contribute to the complexity of the task, limiting overall model performance. Additionally, traditional Euclidean embeddings struggle to capture hierarchical structures and target relationships, reducing expressive power. To address these issues, we propose a Perceptual Consistency Enhancement Module (PCEM) and an Uncertainty Calibration Module (UCM). PCEM ensures consistent feature extraction under dynamic conditions, enhancing robustness to challenging visual contexts.UCM converts features from Euclidean to hypersurface space, dynamically quantifies uncertainty via hypersurface distances, and assigns greater training focus to high-uncertainty regions. This improves both accuracy and robustness. Experiments on OpenTTGames and TableTennis datasets demonstrate substantial improvements in detection accuracy and tracking stability compared to state-of-the-art methods, particularly in complex scenarios.</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

Enhancing perceptual consistency and calibrating uncertainty for real-time table tennis ball detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月30日

查看全文
  • Chang Hong,
  • Xinjie Wang,
  • Shu Zhou

摘要

Driven by advancements in computer vision and machine learning, automated systems have made remarkable strides in capturing real-time dynamics in table tennis. However, table tennis ball detection still faces numerous challenges. First, the ball’s small size, rapid motion, and variable trajectories result in complex image positions, making accurate detection difficult. Moreover, background interferences such as spectators, billboards, and referees can lead to target confusion, while lighting variations further increase detection uncertainty. These factors collectively contribute to the complexity of the task, limiting overall model performance. Additionally, traditional Euclidean embeddings struggle to capture hierarchical structures and target relationships, reducing expressive power. To address these issues, we propose a Perceptual Consistency Enhancement Module (PCEM) and an Uncertainty Calibration Module (UCM). PCEM ensures consistent feature extraction under dynamic conditions, enhancing robustness to challenging visual contexts.UCM converts features from Euclidean to hypersurface space, dynamically quantifies uncertainty via hypersurface distances, and assigns greater training focus to high-uncertainty regions. This improves both accuracy and robustness. Experiments on OpenTTGames and TableTennis datasets demonstrate substantial improvements in detection accuracy and tracking stability compared to state-of-the-art methods, particularly in complex scenarios.