中图分类号：TP391.41 文献标志码：A DOI: 10.16157/j.issn.0258-7998.267775
中文引用格式： 樊泽，李鹏辉，赵皓阳，等. 基于PointMeta和上下文感知的点云场景理解[J]. 电子技术应用，2026，52(9)：149-158.
英文引用格式： Fan Ze，Li Penghui，Zhao Haoyang，et al. Point cloud scene understanding based on PointMeta and context-awareness[J]. Application of Electronic Technique，2026，52(9)：149-158.

Point cloud scene understanding based on PointMeta and context-awareness

Abstract： Aiming at the dual challenges of limited local feature representation and inaccurate semantic alignment in point cloud scene understanding, this paper proposes a 3D understanding framework integrating a hierarchical meta-architecture and dual-stream context-awareness. Firstly, PointMeta is adopted as the visual backbone, and a multi-scale adaptive feature extraction network is constructed through meta-function decomposition to achieve adaptive fusion of multi-scale features. Then, a context-aware spatial-semantic alignment strategy is introduced to align 3D object features and linguistic features from the perspectives of semantics and context, thereby realizing cross-modal information fusion. Additionally, extra spatial information is incorporated into text labels to reduce ambiguities caused by the relationships between contextual features and language. Experiments on scene understanding using public datasets ScanRefer and Nr3D show that the proposed network model achieves significant improvements compared to PointNet++. Compared with the results of other network models on public datasets, the proposed model exhibits varying degrees of performance enhancement, demonstrating stronger generalization ability and favorable practical application value.

Key words : point cloud scene understanding；PointMeta；context-awareness；spatial-semantic alignment