Urban point cloud segmentation improved by adaptive aerial content classifier
Aerial Content-Aware Sample-Adaptive Classifier for Urban Photogrammetric Point Cloud Segmentation
Computer Vision and Pattern Recognition
Summary
Identifying different objects in 3D urban scans can be tricky, especially when objects look similar from close up. The authors address this by combining detailed 3D views with broader overhead aerial images, which show more context but less height info. They designed a new classifier method that adapts to each data sample by connecting these two types of views better, improving accuracy in labeling parts of the scenes. Their approach performed better than previous methods on popular urban 3D datasets.
What this means in practice
- •For urban planners: Automatically classify urban 3D scans more accurately to assist in city modeling and infrastructure assessment.
- •For mapping service providers: Integrate adaptive classifiers to improve the accuracy of feature labeling in aerial and street-level 3D maps.$Commercial implications: Enables production of more detailed and reliable urban maps for commercial GIS and navigation platforms.
Authors
Chong Liu, Yonghua Lu, Zhanglin Cheng
Abstract
Urban photogrammetric point clouds are commonly processed as local 3D patches, limiting the context available for distinguishing geometrically similar categories. Orthographic projection of the same data yields a co-registered digital orthophoto map (DOM)/RGB-D representation that supports efficient large-context modeling but compresses vertical geometry. Existing projection--point methods mainly fuse these two representations at the feature level while retaining a classifier shared across all samples. We present the Aerial Content-Aware Sample-Adaptive Classifier (ACASAC), which complements feature fusion with decision-level adaptation. A point-aligned aerial region bridges the full aerial crop and local 3D patch. ACASAC models scene-to-local variation within the projection and the representation discrepancy between aligned 2D and 3D supports, then transfers both relations to refine the local point class representation. The refined representation, scene-level aerial context, and fixed fused-head prior jointly form a classifier for the current sample. ACASAC improves the fixed dual-representation head by 2.1 and 2.7 mIoU points on SensatUrban and UrbanBIS, respectively, and by 2.1--2.3 points across four tested encoder combinations.