Abstract
As maintaining road networks is labor-intensive, many automatic road extraction approaches have been introduced to solve this real-world problem, fueled by the abundance of large-scale high-resolution satellite imagery and advances in data-driven vision technology. However, their performance is limited to fully automating road map extraction in real-world services. Hence, many services employ the human-in-the-loop approaches on the extracted road maps: semi-automatic detection and repairing faulty road maps. Our paper exclusively focuses on the latter, introducing a novel data-driven approach for fixing road maps. We incorporate image inpainting approaches to tackle complex road geometries without custom-made algorithms for each road shape, yielding a method that is readily applicable to any road map segmentation model. We demonstrate the effectiveness of our method on various real-world road geometries, such as straight and curvy roads, T-junctions, and intersections.
Introduction
Extracting road networks from satellite imagery is essential for modern navigation systems, yet automated extraction remains imperfect. Current systems often produce "broken" maps characterized by missing segments, disconnected intersections, or artifacts caused by environmental occlusions like tree cover and shadows. Traditionally, these errors are addressed through labor-intensive manual annotation or simple heuristic algorithms like curve fitting. However, heuristics often fail to capture the complexity of real-world geometries, such as multi-way intersections or sharp mountain curves.
This research proposes a "human-in-the-loop" approach: once a human or automated tool localizes an error, a specialized neural network "inpaints" the correct road structure based on the surrounding context. By treating faulty road maps as "damaged images" to be repaired, the proposed GLCRC (Globally and Locally Consistent Road map Completion) method moves away from rigid geometric rules toward a flexible, data-driven solution.
Methodology
The GLCRC architecture is specifically designed for the topological constraints of road networks.
Architecture and Training
Based on the Globally and Locally Consistent Image Completion (GLCIC) framework, the model employs a generator and two discriminators. The generator utilizes deep dilated convolutions to expand its receptive field, allowing it to incorporate distant road segments to ensure global connectivity. The dual-discriminator setup evaluates both the global map consistency and the local patch quality.
Loss Functions
To overcome the limitations of standard pixel-wise comparisons, the authors introduced a combination of loss functions:
- Perceptual Loss: Ensures that the generated features and textures align with realistic road patterns, preventing the "blurring" effect common with Mean Square Error (MSE) loss.
- RaLSGAN Loss: A Relativistic Least Square GAN loss is used to produce sharp, binary-like road segments rather than faded artifacts.
Dataset
The model was trained on the Massachusetts Road Geometry dataset, covering 2,600 square kilometers. The high-resolution imagery was processed into 9,972 training and 567 test images (256x256 pixels).
Results and Discussion
The GLCRC model was evaluated against baseline methods using metrics for Correctness, Completeness, and Quality.
| Method | Correctness | Completeness | Quality |
|---|---|---|---|
| Vanilla GLCIC | 0.787 | 0.803 | 0.664 |
| GLCRC+L (Ours) | 0.795 | 0.831 | 0.671 |
The model showed significant improvements in specific geometries, with straight road correctness increasing from 0.787 to 0.894 and T-junction correctness rising from 0.775 to 0.842.
Key Insights
- Geometry vs. Pixels: Unlike standard inpainting, road completion must be "geometrically sensible." The model successfully learns to predict exit points based on entry angles.
- Stability over Creativity: While diffusion models like RePaint were tested, they were found to be too slow and prone to generating "creative" but incorrect paths. The GAN-based approach provided the necessary stability and speed for GIS workflows.
- The Blur Problem: The transition from MSE to perceptual and GAN-based losses was critical. MSE tends to produce "safe" blurry averages, whereas the proposed losses force the model to commit to sharp, usable road structures.
Conclusion
The GLCRC approach provides a robust, data-driven method for repairing faulty road maps. By leveraging inpainting techniques, it handles complex geometries that baffle traditional heuristics. Future work will focus on automating the localization of errors and extending the model to handle multi-lane roads and varied road classes, potentially integrating raw satellite imagery to improve performance under heavy canopy cover.
Summary updated