EUVIMEDEUVIMED entdecken
Crossref

SGM-Net: 3D Point Cloud Class-Incremental Segmentation via Semantic-Aware Global Modeling

Jinshuo Liu, Bingtao Ma, Zhidong Zhao, Chenggang Yan, Shuai Wang

IEEE Signal Processing Letters · 2026 · 33 · 2929-2933

DOI: 10.1109/lsp.2026.3706433 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Jinshuo Liu und 4 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Gefördert von: Zhejiang Provincial Natural Science Foundation of China, National Natural Science Foundation of China, Zhejiang Provincial Key Research and Development Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Point cloud semantic segmentation is a fundamental task in environmental perception. In open-world scenarios, however, it faces major challenges due to the continuous emergence of novel classes. Existing incremental point cloud category learning methods often suffer from catastrophic forgetting and semantic shift. This issue is further exacerbated by the unstructured nature of point clouds and the difficulty in extracting discriminative semantic features. In this paper, we propose the Semantic-Aware Global Modeling Network (SGM-Net) for robust incremental point cloud segmentation. Specifically, we design a Semantic Feature Mamba module that uses Feature Reversal Augmentation to construct an inverted feature sequence from the point sequence, and leverages Mamba to capture long-range dependencies to perform global modeling. In addition, we introduce a Category Token Injection mechanism that incorporates category-level information to enhance semantic representations, thereby improving the model’s understanding of scene semantics and alleviating catastrophic forgetting in incremental learning. Furthermore, we propose a Dual-guided Pseudo-label Generation strategy that utilize uncertainty estimation and feature prototype distance denoising to mitigate confirmation bias. Extensive experiments on the S3DIS dataset demonstrate that SGM-Net outperforms existing methods and achieves a better balance between preserving old knowledge and learning novel classes.

Förderung

  • Zhejiang Provincial Natural Science Foundation of China
  • National Natural Science Foundation of China
  • Zhejiang Provincial Key Research and Development Program

Zitieren

APA

Liu, J., Ma, B., Zhao, Z., Yan, C., & Wang, S. (2026). SGM-Net: 3D Point Cloud Class-Incremental Segmentation via Semantic-Aware Global Modeling. IEEE Signal Processing Letters, 33, 2929-2933. https://doi.org/10.1109/lsp.2026.3706433

Vancouver

Liu J, Ma B, Zhao Z, Yan C, Wang S. SGM-Net: 3D Point Cloud Class-Incremental Segmentation via Semantic-Aware Global Modeling. IEEE Signal Processing Letters. 2026;33:2929-2933. doi:10.1109/lsp.2026.3706433

BibTeX
@article{Liu2026,
  author = {Liu, Jinshuo and Ma, Bingtao and Zhao, Zhidong and Yan, Chenggang and Wang, Shuai},
  title = {SGM-Net: 3D Point Cloud Class-Incremental Segmentation via Semantic-Aware Global Modeling},
  journal = {IEEE Signal Processing Letters},
  year = {2026},
  volume = {33},
  pages = {2929--2933},
  doi = {10.1109/lsp.2026.3706433}
}
Kostenlos starten →