Geometry-aware AI dissects Chinese-style buildings from aerial 3D models

Researchers have unveiled a geometry-aware framework that automatically breaks down Chinese-style buildings into their fine-grained components using aerial photogrammetric 3D models, according to a study published in npj Heritage Science. The work targets a persistent obstacle in digital heritage documentation: the elongated, interconnected, and geometrically complex parts of traditional roofs and eaves that resist conventional segmentation and decomposition methods.

The framework combines three linked components. First, a geometry-adaptive segmentation network extracts elongated architectural components from multi-view projective images derived from the aerial survey data. Second, a geometry-preserving semantic transfer strategy moves high-resolution 2D semantic labels onto the reconstructed 3D models, so that the detail captured in the imagery is not lost during projection into three dimensions. Third, a geometry-constrained decomposition method separates interconnected roof components that would otherwise merge into a single undifferentiated mass.

To evaluate the approach, the team constructed a benchmark dataset with annotations of elongated roof components, providing a standardized test bed for future methods. Experiments reported in the study show that the framework improves the extraction of elongated architectural components in both the multi-view images and the reconstructed 3D models, yielding decomposition results that are more structurally complete, topologically consistent, and geometrically accurate.

The authors position the framework and its annotated dataset as a foundation for future research on digital documentation and structural analysis of heritage buildings with similar characteristics. Because Chinese-style architecture features sweeping rooflines, layered eaves, and interlocking members that are difficult to isolate automatically, tools that respect the underlying geometry during both segmentation and 3D semantic transfer could streamline the creation of detailed digital records of such structures.

The study, published as an open-access article with a permanent DOI, was shared early to provide faster access to peer-reviewed, accepted research; the released version is subject to further edits and will be replaced automatically by the final Version of Record. As with any methodological advance demonstrated on a specific benchmark, performance across other building traditions and survey conditions remains a subject for future evaluation rather than an established result.

Subject of Research: Geometry-aware fine-grained decomposition of Chinese-style buildings from aerial photogrammetric 3D models

Article Title: Geometry-aware fine-grained decomposition of Chinese-style buildings from aerial photogrammetric 3D models

Article References: Gong, Y., Liu, S., Duan, H., Mao, Z., & Huang, X. (2026). Geometry-aware fine-grained decomposition of Chinese-style buildings from aerial photogrammetric 3D models. npj Heritage Science. https://doi.org/10.1038/s40494-026-03054-w

Image Credits: AI Generated

DOI: 10.1038/s40494-026-03054-w

Keywords: photogrammetry, 3D models, Chinese-style architecture, heritage documentation, semantic segmentation, roof components, geometry-aware framework, digital heritage, benchmark dataset, fine-grained decomposition, Geometry-aware, fine-grained