Gated multi-source fusion with geometric sequence modeling for novel urban structure discovery.
Saved in:
| Title: | Gated multi-source fusion with geometric sequence modeling for novel urban structure discovery. |
|---|---|
| Authors: | Du, Jing1 (AUTHOR), Zelek, John1 (AUTHOR), Zhang, Dedong1 (AUTHOR) zhangdedong@jxl.ac.cn, Li, Jonathan1,2 (AUTHOR) junli@uwaterloo.ca |
| Source: | ISPRS Journal of Photogrammetry & Remote Sensing. Dec2025, Vol. 230, p495-523. 29p. |
| Subject Terms: | *GEOMETRIC modeling, *URBAN morphology, *ARTIFICIAL neural networks, *MULTISENSOR data fusion, *MACHINE learning, *KNOWLEDGE transfer |
| Abstract: | Novel Class Discovery (NCD) in 3D point cloud semantic segmentation presents critical challenges for urban management systems, where models must segment previously unseen object classes in rapidly evolving urban environments. Traditional 3D semantic segmentation models struggle to adapt to heterogeneous spatial characteristics and complex geometric structures of urban point clouds, limiting their ability to handle novel objects without extensive retraining. This paper introduces Adaptive Geometric Discovery Network (AGDNet), a comprehensive framework enhancing NCD through three key innovations: Adaptive Geometric Sequence Modeling module (AGSM), Dynamic Gaussian Embedding module (DGE), and Gated Multi-Source Feature Fusion module (GMSFF). AGSM addresses heterogeneous spatial characteristics through density-aware adaptive sampling, dynamic grouping, and multi-aspect geometric feature encoding. DGE represents point clouds as learnable 3D Gaussians parameterized by position, scale, orientation, and features, providing continuous probabilistic representations capturing both local geometric details and global spatial contexts. GMSFF integrates features from AGSM, DGE, and MinkowskiNet through context-aware gating mechanisms. The framework introduces three specialized knowledge transfer objectives for NCD: Prototype Relation Loss establishes semantic connections between known and novel class prototypes; Contrastive Alignment Loss creates instance-level semantic bridges; and Semantic Transfer Loss enables distribution-based knowledge propagation. These objectives bridge the semantic gap between known and novel categories while mitigating class imbalance challenges. Comprehensive evaluation on Toronto-3D, SemanticSTF, and SemanticPOSS datasets demonstrates significant improvements over state-of-the-art methods NOPS and CHNCD. For novel class discovery, the framework achieves average improvements of 6.47%/3.48%, 4.61%/3.12%, and 6.64%/4.24% in novel class mean Intersection over Union (mIoU) over NOPS/CHNCD respectively. For overall performance, improvements reach 6.59%/3.88%, 7.32%/4.80%, and 7.27%/4.62% in overall mIoU. These results validate the framework's effectiveness for urban management, environmental monitoring, and infrastructure planning applications. [ABSTRACT FROM AUTHOR] |
| Database: | Academic Search Index |
Full Text Finder 
Nájsť tento článok vo Web of Science | Abstract: | Novel Class Discovery (NCD) in 3D point cloud semantic segmentation presents critical challenges for urban management systems, where models must segment previously unseen object classes in rapidly evolving urban environments. Traditional 3D semantic segmentation models struggle to adapt to heterogeneous spatial characteristics and complex geometric structures of urban point clouds, limiting their ability to handle novel objects without extensive retraining. This paper introduces Adaptive Geometric Discovery Network (AGDNet), a comprehensive framework enhancing NCD through three key innovations: Adaptive Geometric Sequence Modeling module (AGSM), Dynamic Gaussian Embedding module (DGE), and Gated Multi-Source Feature Fusion module (GMSFF). AGSM addresses heterogeneous spatial characteristics through density-aware adaptive sampling, dynamic grouping, and multi-aspect geometric feature encoding. DGE represents point clouds as learnable 3D Gaussians parameterized by position, scale, orientation, and features, providing continuous probabilistic representations capturing both local geometric details and global spatial contexts. GMSFF integrates features from AGSM, DGE, and MinkowskiNet through context-aware gating mechanisms. The framework introduces three specialized knowledge transfer objectives for NCD: Prototype Relation Loss establishes semantic connections between known and novel class prototypes; Contrastive Alignment Loss creates instance-level semantic bridges; and Semantic Transfer Loss enables distribution-based knowledge propagation. These objectives bridge the semantic gap between known and novel categories while mitigating class imbalance challenges. Comprehensive evaluation on Toronto-3D, SemanticSTF, and SemanticPOSS datasets demonstrates significant improvements over state-of-the-art methods NOPS and CHNCD. For novel class discovery, the framework achieves average improvements of 6.47%/3.48%, 4.61%/3.12%, and 6.64%/4.24% in novel class mean Intersection over Union (mIoU) over NOPS/CHNCD respectively. For overall performance, improvements reach 6.59%/3.88%, 7.32%/4.80%, and 7.27%/4.62% in overall mIoU. These results validate the framework's effectiveness for urban management, environmental monitoring, and infrastructure planning applications. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 09242716 |
| DOI: | 10.1016/j.isprsjprs.2025.09.017 |