Scholarlink

<p>Semantic segmentation of point clouds is challenged by their unstructured nature, data sparsity, and the prohibitive cost of manual annotations, which are further exacerbated in complex indoor environments with fine-grained object arrangements. To address these issues, we propose CDHC3D, a novel unsupervised segmentation framework that effectively bridges 2D–3D modality gaps and enhances semantic consistency through a two-stage learning paradigm. In the cross-modal distillation stage, we introduce a dual modality adaptive fusion module that aligns 2D texture features with 3D geometric data and performs two-stage voxel-level feature distillation, thereby mitigating modal heterogeneity and enhancing feature discriminability without supervision. In the hierarchical clustering stage, we design a multi-scale supervoxel-based clustering strategy with local neighborhood enhancement and cross-layer semantic propagation, enabling the model to progressively refine pseudo-labels and capture both global structure and local details. Extensive experiments on ScanNet-v2 and S3DIS show that CDHC3D achieves competitive performance and offers consistent improvements over recent unsupervised methods in both segmentation accuracy and structural consistency. Our results highlight the potential of combining cross-modal knowledge transfer and hierarchical pseudo-label refinement for label-free 3D scene understanding.</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

CDHC3D: unsupervised indoor 3D semantic segmentation via cross-modal distillation and hierarchical clustering

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Linxian Zhu,
  • Shengjie Ren,
  • Chengyi Zhang,
  • Miaohui Zhang

摘要

Semantic segmentation of point clouds is challenged by their unstructured nature, data sparsity, and the prohibitive cost of manual annotations, which are further exacerbated in complex indoor environments with fine-grained object arrangements. To address these issues, we propose CDHC3D, a novel unsupervised segmentation framework that effectively bridges 2D–3D modality gaps and enhances semantic consistency through a two-stage learning paradigm. In the cross-modal distillation stage, we introduce a dual modality adaptive fusion module that aligns 2D texture features with 3D geometric data and performs two-stage voxel-level feature distillation, thereby mitigating modal heterogeneity and enhancing feature discriminability without supervision. In the hierarchical clustering stage, we design a multi-scale supervoxel-based clustering strategy with local neighborhood enhancement and cross-layer semantic propagation, enabling the model to progressively refine pseudo-labels and capture both global structure and local details. Extensive experiments on ScanNet-v2 and S3DIS show that CDHC3D achieves competitive performance and offers consistent improvements over recent unsupervised methods in both segmentation accuracy and structural consistency. Our results highlight the potential of combining cross-modal knowledge transfer and hierarchical pseudo-label refinement for label-free 3D scene understanding.