Scholarlink

<p>In 3D point cloud analysis, effectively bridging local geometric details and global semantics, and alleviating feature bias caused by non-uniform point cloud distribution are key challenges. Although existing Transformer methods have long-range modeling capability, they often ignore local structured semantics and do not sufficiently consider the differences in feature dependencies required by different tasks. To this end, this paper proposes a Dynamic Part-Center-based Point Cloud Transformer (DPCT). First, the Part-Center mechanism is designed to enhance the model’s representation and utilization of local structural semantics. Second, the Task-modulated Self-positioned Global Feature Modeling module is proposed, which uses the part-center distribution to modulate global anchors and formulates corresponding feature dependency strategies for classification and segmentation tasks. A density-aware module is also introduced to alleviate the influence of non-uniform point cloud distribution. In addition, the Global Semantic Guidance mechanism is designed for the part segmentation task to enhance the consistency between local predictions and the overall structure. Experiments are conducted on three standard datasets, ScanObjectNN, ModelNet40, and ShapeNetPart. The experimental results show that DPCT achieves good performance in point cloud classification and object-level part segmentation tasks. In particular, in the ShapeNetPart part segmentation task, DPCT achieves competitive results among the methods included for comparison in this paper, with Instance mIoU and Class mIoU of 87.2% and 85.5%, respectively.</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

DPCT: dynamic part-center-based point cloud transformer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月27日

查看全文
  • Qiang Sang,
  • Yunxiao Zhang

摘要

In 3D point cloud analysis, effectively bridging local geometric details and global semantics, and alleviating feature bias caused by non-uniform point cloud distribution are key challenges. Although existing Transformer methods have long-range modeling capability, they often ignore local structured semantics and do not sufficiently consider the differences in feature dependencies required by different tasks. To this end, this paper proposes a Dynamic Part-Center-based Point Cloud Transformer (DPCT). First, the Part-Center mechanism is designed to enhance the model’s representation and utilization of local structural semantics. Second, the Task-modulated Self-positioned Global Feature Modeling module is proposed, which uses the part-center distribution to modulate global anchors and formulates corresponding feature dependency strategies for classification and segmentation tasks. A density-aware module is also introduced to alleviate the influence of non-uniform point cloud distribution. In addition, the Global Semantic Guidance mechanism is designed for the part segmentation task to enhance the consistency between local predictions and the overall structure. Experiments are conducted on three standard datasets, ScanObjectNN, ModelNet40, and ShapeNetPart. The experimental results show that DPCT achieves good performance in point cloud classification and object-level part segmentation tasks. In particular, in the ShapeNetPart part segmentation task, DPCT achieves competitive results among the methods included for comparison in this paper, with Instance mIoU and Class mIoU of 87.2% and 85.5%, respectively.