Segmenting individual roof planes from 3D point clouds is a foundational task for automated 3D city modeling and planning rooftop solar installations. While Transformer-based models have revolutionized point cloud processing, they often struggle with the geometric rigors of roof structures. SPPSFormer (SuperPoint Plane Segmentation transFormer) addresses these challenges by combining high-quality superpoint generation, Kolmogorov-Arnold Networks (KAN), and strategic geometric post-processing.
The Problem: Geometric Precision vs. Data Noise
Existing deep learning models for roof segmentation face two primary hurdles. First, standard backbones often fail to capture the global context required for complex roof shapes. Second, "superpoint" Transformers—which group points into clusters to reduce computation—frequently utilize low-quality clusters that blur boundaries between adjacent planes.
Furthermore, high-fidelity point-level annotations for roof boundaries are scarce and often contain errors. SPPSFormer is designed to be geometrically precise while remaining robust to such imperfect training data.
The SPPSFormer Architecture
The framework utilizes a 3D UNet backbone to extract features, which are then aggregated based on generated superpoints. These features pass through a Transformer decoder and a specialized prediction head that incorporates Fourier-based Kolmogorov-Arnold Networks.
Key Methodological Pillars
The SPPSFormer architecture is built on four innovative components:
- Two-Stage Superpoint Generation: Instead of simple clustering, a two-stage approach is used. Stage 1 ensures boundary accuracy via region growing, while Stage 2 ensures uniformity using K-means. This consistency stabilizes Transformer learning.
- Handcrafted Geometric Features: Beyond XYZ coordinates, the model incorporates multidimensional features like linearity, planarity, and a novel "contour" feature to improve performance on limited data.
- Hybrid Decoder (Fourier KAN): Standard Multi-Layer Perceptrons (MLPs) in the mask branch are replaced with Fourier-based Kolmogorov-Arnold Networks (Fourier KAN). KANs approximate complex geometric functions more efficiently than traditional MLPs.
- Geometric Post-processing: A plane completion module identifies missed planes, and a boundary refinement module sharpens edges. This allows the neural network to focus on the "main body" of planes while traditional geometry handles the precision.
Experimental Evaluation and Results
The model was evaluated on the RoofN3D dataset and a newly annotated Building3D benchmark containing 10,539 buildings.
- State-of-the-Art Performance: SPPSFormer outperformed previous methods like DeepRoofPlane and SPFormer. On RoofN3D, it improved Coverage (Cov) by 3.03% and Weighted Coverage (WCov) by 1.85%.
- Robustness to Noise: In "stress tests" with corrupted boundary labels, SPPSFormer maintained performance where other models failed, thanks to its geometric refinement stage.
- Efficiency: A "nano" version with 90% fewer parameters (1.8M) still outperformed many full-sized competitors.
Key Insights
The researchers found that superpoint quality is a critical bottleneck; "mixed" superpoints containing multiple instances propagate errors that deep learning alone cannot resolve. Additionally, the success of KANs suggests they are particularly well-suited for the continuous, smooth functions inherent in geometric point cloud data, outperforming standard MLPs in mask prediction.
Conclusion
SPPSFormer demonstrates that the path to high-fidelity 3D segmentation lies in the synergy between modern Transformer architectures and traditional geometric principles. By prioritizing superpoint quality and boundary refinement, the model achieves superior accuracy and robustness, setting a new standard for roof plane instance segmentation.
Summary updated