Spatial relationships guide semantic component extraction from Gaussian splatting of traditional Chinese timber architecture
Three-dimensional Gaussian splatting enables photorealistic documentation of architectural heritage, but extracting semantically accurate components remains challenging. We propose a framework for traditional Chinese timber architecture that combines spatial relationship-guided annotation with semantic–appearance decoupled Gaussian extraction. A spatial relationship graph guides a multimodal large language model in distinguishing visually similar components. Semantic assignments and filtering weights are then learned while keeping the reconstructed scene parameters fixed. On a Qing-dynasty timber-frame building with 22 component categories, the framework increases image-level mean intersection over union from 69.03% to 85.72% and extracted-object mean intersection over union from 28.01% to 69.24% compared with Heritage-3DGS. Average isolated-component peak signal-to-noise ratio increases from 17.05 to 24.30 dB, while full-scene rendering quality is preserved. The framework provides independently renderable component categories for semantic access to photorealistic architectural heritage records.
Authors
- Dong Li (ORCID: https://orcid.org/0000-0003-1364-9768)
- Xingxing Xiao (ORCID: https://orcid.org/0000-0003-1045-6880)
- Weifeng Ma
- Ming Huang
- Lei Wang
Institutions
- Yunnan Normal University (CN)
- Beijing University of Civil Engineering and Architecture (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11518276/v1
- Primary Topic
- 3D Surveying and Cultural Heritage
- Type
- preprint
Funders
- National Natural Science Foundation of China
- National Science and Technology Major Project