An aerial point cloud classification using point transformer via multi-feature fusion

Abstract Point Transformer can effectively capture both local information and long-range global context in point cloud data. However, its reliance on dividing local areas into separate tokens can result in a loss of instance structure, limiting its feature representation ability in large-scale urban...

Full description

Saved in:
Bibliographic Details
Main Authors: Jiechen Pan, Jiayin Cao, Shuai Xing, Mofan Dai, Jikun Liu, Xuying Wang, Yunsheng Zhang, Gaoshuang Huang
Format: Article
Language:English
Published: Nature Portfolio 2025-07-01
Series:Scientific Reports
Subjects:
Online Access:https://doi.org/10.1038/s41598-025-02719-z
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract Point Transformer can effectively capture both local information and long-range global context in point cloud data. However, its reliance on dividing local areas into separate tokens can result in a loss of instance structure, limiting its feature representation ability in large-scale urban aerial point clouds, especially the fine-grained classification of ground objects. To address these limitations, we propose a novel Point Transformer-based Multi-feature Fusion (PTMF) Network that explicitly integrates geometric features into the Point Transformer architecture to classify large-scale aerial point clouds. The fusion of geometric features effectively complements global contextual feature extraction that is solely based on positional relationships with prior information. Specifically, the PTMF network fuses down-sampled geometric and inherent point cloud features at each stage, using the Transition Up module to up-sample mapping features. Our experiments on the SensatUrban and DALES datasets demonstrate significant improvements, achieving mIoU scores of 63.52% and 82.18%, respectively, outperforming existing methods.
ISSN:2045-2322