1. Hierarchical 2-D/3-D Object-Based Classification of Photogrammetric Textured Mesh Models
- Author
-
Zhongwen Hu, Jinhua Zhang, Zhigang Liu, Yinghui Zhang, Jingzhe Wang, Qian Zhang, and Guofeng Wu
- Subjects
Hierarchical classification ,object-based approach ,random forest (RF) algorithm ,stereoscopic hierarchy ,textured mesh model (TMM) ,Ocean engineering ,TC1501-1800 ,Geophysics. Cosmic physics ,QC801-809 - Abstract
The photogrammetric 3-D textured mesh model (TMM) obtained by unmanned aerial vehicle provides accurate geometric shapes and realistic textures. The 3-D semantics derived from TMMs serve as the foundation in many applications, such as urban planning, forestry, and smart city. Although 3-D TMM provides more features than 2-D images, current classification methods have not fully utilized these features, particularly the stereoscopic hierarchical structure of different objects. To address this issue, we propose a hierarchical object-based method for the classification of TMMs, consisting of three key steps: 1) the TMM is first hierarchically segmented into ground surface meshes and off-ground 3-D objects using a cloth-simulated filtering algorithm; 2) the ground surface mesh is projected to 2-D ortho-image, where object-based image classification is used to classify pixels. The resulting semantic labels of pixels are then mapped back to the corresponding mesh model; 3) instance-level 3-D objects are created through connected component analysis of off-ground meshes, and then classified via a 3-D object-based approach. The 3-D semantic model is generated by merging the outcomes of steps 2) and 3). Experimental results indicate that the stereoscopic hierarchical strategy effectively decomposes a 3-D scene into simple ground surfaces and off-ground 3-D objects, yielding improvements in both accuracy and efficiency. Our proposed method demonstrates an accuracy increase of over 7%–36% compared to existing methods. This is the first time a stereoscopic hierarchy has been introduced for classifying 3-D textured mesh model, providing a valuable reference for future classification methods.
- Published
- 2025
- Full Text
- View/download PDF