Abstract:The photogrammetric three-dimensional (3D) textured mesh model (TMM) obtained by unmanned aerial vehicle provides accurate geometric shapes and realistic textures. The 3D semantics derived from TMMs serves as the foundation in many applications, such as urban planning, forestry, and smart city. Although 3D TMM provide more features than 2D 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 firstly hierarchically segmented into ground surface meshes and off-ground 3D objects using a cloth-simulated filtering algorithm. 2) The ground surface mesh is projected to 2D 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 3D objects are created through connected component analysis of off-ground meshes, and then classified via a 3D object-based approach. The 3D semantic model is generated by merging the outcomes of steps (2) and (3). Experimental results indicate that the stereoscopic hierarchical strategy effectively decomposes a 3D scene into simple ground surfaces and off-ground 3D 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 3D textured mesh model, providing a valuable reference for future classification methods.

Volumetric Part Based 3D Object Classification

3d Object Classification Based On Volumetric Parts

3D object classification system based on volumetric parts

3d Object Classification By Part Features Fusion

Semantic-oriented 3D model classification and retrieval using Gaussian processes

Extracting Superquadric-based Geon Description for 3D Object Recognition

3D Object Classification Based on Improved Interpretation Tree

Superquadric Similarity Measure With Spherical Harmonics In 3d Object Recognition

Unify 3D Shape Retrieval and Classification in One Framework

Part based Structural Description Extracting and Modeling

A Unified Feature Representation and Learning Framework for 3D Shape

3d Model Classification Based On Multiple Features Integration

PartGS:Learning Part-aware 3D Representations by Fusing 2D Gaussians and Superquadrics

Voxel-based three-view hybrid parallel network for 3D object classification

Hierarchical 2D/3D Object-based Classification of Photogrammetric Textured Mesh Models

Learning Part-aware 3D Representations by Fusing 2D Gaussians and Superquadrics

3x2: 3D Object Part Segmentation by 2D Semantic Correspondences

Superquadric-based geons recognition utilizing support vector machines

A Novel 3d Cad Model Retrieval Method Based on Vertices Classification and Weights Combination Optimization

3D Model Classification Based on GCN and SVM

Study On Similarity Measure Of Superquadrics