Scholarlink

<p>3D indoor scene perception based on point clouds has received increasing attention due to its ability to represent a large amount of information with less storage space. In order to solve the problem that graphical information of local features is often overlooked, we propose a multi-scale graph convolution network (MGCNet) with an encoder-decoder framework utilizing a hierarchical structure and local area feature extraction. In the encoder, a multi-scale graph convolutional feature learning block, which includes a graph convolution module and a multi-scale feature aggregation (MFA) module, is constructed to increase the learning capability of local graphic features. The Kolmogorov-Arnold Networks (KAN) module integrated into the decoder can address complex nonlinear problems by effectively capturing global context information, allowing the model to accurately extract target features while suppressing interference from complex background noise. Extensive experiments on public datasets show that MGCNet outperforms other comparative methods for indoor scene segmentation.</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

Multi-scale local geometry feature and global context learning with Kolmogorov-Arnold representation for 3D semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Fan Zhang,
  • Huazhen Zhang,
  • Yun Wang

摘要

3D indoor scene perception based on point clouds has received increasing attention due to its ability to represent a large amount of information with less storage space. In order to solve the problem that graphical information of local features is often overlooked, we propose a multi-scale graph convolution network (MGCNet) with an encoder-decoder framework utilizing a hierarchical structure and local area feature extraction. In the encoder, a multi-scale graph convolutional feature learning block, which includes a graph convolution module and a multi-scale feature aggregation (MFA) module, is constructed to increase the learning capability of local graphic features. The Kolmogorov-Arnold Networks (KAN) module integrated into the decoder can address complex nonlinear problems by effectively capturing global context information, allowing the model to accurately extract target features while suppressing interference from complex background noise. Extensive experiments on public datasets show that MGCNet outperforms other comparative methods for indoor scene segmentation.