Abstract
SCENE-Net is a "white-box" 3D semantic segmentation model designed for identifying power line towers in large-scale, low-resource LiDAR datasets. By utilizing Group Equivariant Non-Expansive Operators (GENEOs) and just 11 trainable geometric parameters, the model achieves state-of-the-art efficiency and intrinsic interpretability. It trains in under 90 minutes on a standard laptop and demonstrates high robustness to noisy labeling and class imbalance, outperforming traditional "black-box" CNNs in precision and parameter efficiency.
Introduction
Deep learning has revolutionized 3D semantic segmentation, yet state-of-the-art models often function as "black boxes" requiring millions of parameters and massive computational resources. For industries like electrical utilities, which must inspect thousands of kilometers of power lines using helicopter-based LiDAR scans, these requirements are prohibitive. Such datasets are typically "low-resource": they are highly imbalanced, contain noisy labels, and must be processed without supercomputing clusters. Furthermore, infrastructure maintenance requires transparency—engineers must understand why a model flags a specific object.
Methodology: GENEOs and Signature Shapes
The authors propose SCENE-Net, a model that leverages Group Equivariant Non-Expansive Operators (GENEOs) instead of traditional convolutional neural networks (CNNs). This approach shifts the focus from data brute-force to knowledge engineering.
Key technical components include:
- Geometric Priors: The model uses "Signature Shapes"—mathematically defined kernels representing a Cylinder (tower body), an Arrow (tower and power lines), and a Negative Sphere (to penalize tree-like clusters).
- Equivariance: Operators are designed to be equivariant, ensuring the model recognizes a tower regardless of its rotation or translation in 3D space.
- White-Box Optimization: SCENE-Net learns only 11 parameters corresponding to physical dimensions (e.g., radius, height) and "mixing weights" for the final decision.
- Convex Combination: The final observer is a convex combination of these operators, ensuring mathematical stability.
Datasets
The study utilizes two primary datasets:
- TS40K: A new dataset covering 40,000 km of rural terrain with over 9 billion points. It features extreme class imbalance (towers comprise only 0.52% of points) and noisy ground truth.
- SemanticKITTI: A standard benchmark for autonomous driving used to validate performance against established "black-box" models.
Key Results and Discussion
SCENE-Net demonstrates remarkable efficiency and robustness compared to traditional architectures.
- Extreme Efficiency: With only 11 trainable parameters, SCENE-Net is orders of magnitude smaller than baseline CNNs (2,190 parameters) or SOTA models like SPVNAS (12.5 million parameters).
- Hardware Agnostic: The model trains in approximately 85 minutes on a standard laptop with an inference time of 20ms.
- Performance and Robustness: On the TS40K dataset, SCENE-Net achieved a Precision of 0.82, significantly outperforming the baseline CNN (0.44). Its geometric constraints make it resistant to noisy labels; for instance, it ignores mislabeled flat ground because it does not match the "cylinder" or "arrow" profiles.
On the SemanticKITTI "pole" segmentation task, SCENE-Net achieved an IoU of 57.5, remaining competitive with massive SOTA models while maintaining a parameter efficiency score nearly 5x higher than its closest competitor.
Conclusion
SCENE-Net proves that incorporating domain knowledge through geometric signatures allows for highly efficient, interpretable, and robust 3D segmentation. While currently requiring manual definition of "signature shapes," future work aims to develop a broader library of shapes to detect more complex, organic structures and expand to multi-class segmentation.
Summary updated