Scholarlink

<p>Point cloud segmentation is fundamental to 3D visual understanding but remains challenging due to the unordered and sparse nature of point data. Existing lightweight architectures such as PointNeXt-S aggregate local features with equal channel contributions and lack mechanisms to suppress geometrically non-salient points, limiting fine-grained local modeling. Moreover, standard training configurations struggle with instance-level variations in density and geometry. To address these limitations, we propose GE-PointNeXt, which enhances the Set Abstraction (SA) module with a unified Gating and ECA-Enhanced SA (GE-SA) block. To model inter-channel dependencies and amplify informative channels, we design a lightweight Efficient Channel Attention (ECA) module adapted to the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(C \times N \times K\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>C</mi> <mo>×</mo> <mi>N</mi> <mo>×</mo> <mi>K</mi> </mrow> </math></EquationSource> </InlineEquation> structure of local point neighborhoods. To suppress noisy and non-salient points, we present a self-gating mechanism that modulates each point’s feature by its own activation strength. To improve training stability, we combine Instance–Batch Normalization (IBN) with a hybrid cosine–multi-step learning rate scheduler (CM-LR). Experiments on ShapeNetPart and S3DIS demonstrate that GE-PointNeXt achieves 83.1% class mIoU and 85.7% instance mIoU on ShapeNetPart, and 64.4% mIoU on S3DIS Area&#xa0;5, outperforming the PointNeXt-S baseline and several state-of-the-art methods without increasing model parameters.</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

GE-PointNeXt: self-gating and ECA-Enhanced set abstraction with hybrid training strategies for point cloud segmentation

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Xianglong Xie,
  • Xifan Yao

摘要

Point cloud segmentation is fundamental to 3D visual understanding but remains challenging due to the unordered and sparse nature of point data. Existing lightweight architectures such as PointNeXt-S aggregate local features with equal channel contributions and lack mechanisms to suppress geometrically non-salient points, limiting fine-grained local modeling. Moreover, standard training configurations struggle with instance-level variations in density and geometry. To address these limitations, we propose GE-PointNeXt, which enhances the Set Abstraction (SA) module with a unified Gating and ECA-Enhanced SA (GE-SA) block. To model inter-channel dependencies and amplify informative channels, we design a lightweight Efficient Channel Attention (ECA) module adapted to the \(C \times N \times K\) C × N × K structure of local point neighborhoods. To suppress noisy and non-salient points, we present a self-gating mechanism that modulates each point’s feature by its own activation strength. To improve training stability, we combine Instance–Batch Normalization (IBN) with a hybrid cosine–multi-step learning rate scheduler (CM-LR). Experiments on ShapeNetPart and S3DIS demonstrate that GE-PointNeXt achieves 83.1% class mIoU and 85.7% instance mIoU on ShapeNetPart, and 64.4% mIoU on S3DIS Area 5, outperforming the PointNeXt-S baseline and several state-of-the-art methods without increasing model parameters.