Scholarlink

With the development of fields such as autonomous driving, smart factories, and augmented reality, the demand for scene understanding and environmental perception of point clouds is increasing. Point cloud semantic segmentation technology is a key research focus in this area. However, in large-scale point cloud scenes, some objects' spatial geometric structures do not differ significantly from each other, and relying solely on spatial geometric information can lead to errors in segmentation boundaries. Therefore, this paper synchronizes and encodes the spatial position and color information of point clouds, combined with semantic information, proposing a method of multidimensional local information re-encoding, which enhances the accuracy of segmentation in areas where geometric structures are not significantly different. Subsequently, a residual connection and dense connection module were designed to alleviate the problem of segmentation errors at the boundaries of different categories of objects. Finally, the network performance was evaluated on the S3DIS and Toronto-3D datasets, demonstrating higher overall accuracy compared to PointNet++, RandLA-NET, and BAF-LAC.

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

MDLIRE-Net: Semantic Segmentation of Point Clouds Based on Multidimensional Local Information Re-encoding

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Shijian Huang,
  • Qi Wang,
  • Huasong Min

摘要

With the development of fields such as autonomous driving, smart factories, and augmented reality, the demand for scene understanding and environmental perception of point clouds is increasing. Point cloud semantic segmentation technology is a key research focus in this area. However, in large-scale point cloud scenes, some objects' spatial geometric structures do not differ significantly from each other, and relying solely on spatial geometric information can lead to errors in segmentation boundaries. Therefore, this paper synchronizes and encodes the spatial position and color information of point clouds, combined with semantic information, proposing a method of multidimensional local information re-encoding, which enhances the accuracy of segmentation in areas where geometric structures are not significantly different. Subsequently, a residual connection and dense connection module were designed to alleviate the problem of segmentation errors at the boundaries of different categories of objects. Finally, the network performance was evaluated on the S3DIS and Toronto-3D datasets, demonstrating higher overall accuracy compared to PointNet++, RandLA-NET, and BAF-LAC.