Overview
The UnderOneFacade dataset is presented as a benchmark for 3D facade semantic segmentation. It focuses on the requirements for globally consistent semantic digital twins, specifically the need for centimeter-accurate and geographically transferable 3D facade segmentation. This benchmark comprises centimeter-accurate point clouds with hierarchical, harmonized, and architecturally grounded semantic labels.
Research Context
The development of globally consistent semantic digital twins necessitates precise and geographically transferable 3D facade segmentation. Current progress in facade parsing has limitations, primarily due to an absence of large-scale, standardized benchmarks for evaluating cross-domain generalization. Existing datasets are characterized by their geographically narrow scope, inconsistent semantics, or insufficient precision. This limits the ability to develop models that can generalize across diverse architectural styles and urban environments worldwide.
Approach
The UnderOneFacade benchmark is described as the largest cross-country and cross-continent 3D facade benchmark available. It encompasses 2.7 billion annotated points. The annotations provide hierarchical, harmonized, and architecturally grounded semantic labels. The dataset's design aims to combine geometric precision with standardized semantics at an unprecedented scale. To assess the capabilities of existing methodologies, a systematic evaluation was conducted using representative point-, graph-, and transformer-based architectures against the UnderOneFacade dataset. The evaluation aimed to determine how well these methods perform in recognizing fine-grained architectural elements and their transferability across different geographic domains.
Findings
- UnderOneFacade constitutes a dataset of 2.7 billion annotated points, providing centimeter-accurate point clouds with hierarchical, harmonized, and architecturally grounded semantic labels.
- Systematic evaluation of representative point-, graph-, and transformer-based architectures indicates that current methods encounter difficulties in recognizing fine-grained architectural elements.
- The performance of existing models degrades significantly when applied across different geographic domains.
- The best-performing models achieved an Intersection over Union (IoU) of up to 33 on the fine-grained LoFG3 benchmark, suggesting a substantial gap in current model capabilities for robust and transferable 3D segmentation.
Why This Matters
The UnderOneFacade dataset establishes a rigorous benchmark for developing robust and transferable 3D segmentation models. This is critical for advancing the creation of globally consistent semantic digital twins, which demand centimeter-accurate and geographically transferable 3D facade segmentation. The observed struggles of current methods with fine-grained elements and cross-domain generalization highlight the need for further research and development in this area.