Scholarlink

<p>To address the scarcity of publicly available underground coal mine monitoring image datasets and the challenges of accurate recognition under complex drilling environments, this study constructs a large-scale image dataset for coal mine drill pipe counting based on Oriented Bounding Boxes (CMDPC_OBB). The dataset is designed for multi-pose and multi-view drill pipe detection and counting tasks, improving data coverage in complex underground scenarios. CMDPC_OBB consists of two sub-datasets: a multi-object detection dataset (MOD_2D) and a structurally enhanced single-object classification dataset (SOC_3D). MOD_2D is built using a 15-frame interval sampling strategy, resulting in 114,869 field images annotated with rotated bounding boxes. SOC_3D extends the original samples at the data level by generating eight additional views per instance through single-image 3D reconstruction, yielding 4,023 images to enhance multi-view representation. Nine object detection and oriented detection models were evaluated on CMDPC_OBB. The highest mAP reaches 89.1%, demonstrating the dataset’s effectiveness and benchmarking value in complex underground environments.</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

CMDPC_OBB: A Large-Scale Image Dataset for Coal Mine Drill Pipe Counting based on Oriented Bounding Box

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Fukai Zhang,
  • Xiaoran Liu,
  • Haiyan Zhang,
  • Feng Guo,
  • Shan Zhao,
  • Yanmei Zhang,
  • Guan Yuan,
  • Lu Dong,
  • Xu Chen,
  • Zhanqiang Huo

摘要

To address the scarcity of publicly available underground coal mine monitoring image datasets and the challenges of accurate recognition under complex drilling environments, this study constructs a large-scale image dataset for coal mine drill pipe counting based on Oriented Bounding Boxes (CMDPC_OBB). The dataset is designed for multi-pose and multi-view drill pipe detection and counting tasks, improving data coverage in complex underground scenarios. CMDPC_OBB consists of two sub-datasets: a multi-object detection dataset (MOD_2D) and a structurally enhanced single-object classification dataset (SOC_3D). MOD_2D is built using a 15-frame interval sampling strategy, resulting in 114,869 field images annotated with rotated bounding boxes. SOC_3D extends the original samples at the data level by generating eight additional views per instance through single-image 3D reconstruction, yielding 4,023 images to enhance multi-view representation. Nine object detection and oriented detection models were evaluated on CMDPC_OBB. The highest mAP reaches 89.1%, demonstrating the dataset’s effectiveness and benchmarking value in complex underground environments.