TL;DR
This paper introduces the Global Collinearity-aware Polygonizer (GCP), a novel method for converting building masks from remote sensing images into precise polygonal footprints. GCP refines initial contours and then simplifies them using a dynamic programming approach that prioritizes collinearity, leading to more accurate and compact building representations.
What Problem Does This Paper Address?
Extracting building footprints from high-resolution remote sensing images is crucial for urban planning and analysis. While many methods can generate building footprints as raster-based binary masks, there's a significant need for vector-based polygonal representations. These vector formats are more compact and structured, enabling applications like 3D building modeling. Existing mask-to-polygon methods often struggle with inaccuracies at building corners due to the discrete nature of masks, and traditional simplification algorithms can lead to performance degradation or overly complex polygons. This paper aims to bridge this gap by providing a robust and accurate method for polygonal building mapping directly from binary masks.
Methods
The proposed Global Collinearity-aware Polygonizer (GCP) operates in three main stages:
- Initial Contour Generation: This stage starts with binary masks generated by an instance segmentation model (specifically, Mask2Former is used). These masks are converted into initial polygons using contour tracing. To handle varying densities, these polylines are uniformly sampled and then segmented into fixed-length sequences using a sliding window approach.
- Polyline Regression: A transformer-based regression head is employed to refine these initial, often coarse, polylines. This module takes the sampled polylines and associated image features to predict an offset, resulting in a more accurate representation of the building contours.
- Global Collinearity-aware Polygon Simplification: This is the core innovation. Instead of relying on local simplification methods like Douglas-Peucker, GCP uses a dynamic programming approach. It optimizes an objective function that balances two factors:
- Distance Preservation: Minimizing the distance between the original points and the segments of the simplified polygon.
- Collinearity Preservation: Penalizing the inclusion of unnecessary vertices that do not contribute to capturing the inherent straightness of building edges. This objective is designed to find a globally optimal simplified polyline that is both accurate and compact. The method also incorporates a global collinearity loss to supervise the entire pipeline end-to-end.
Datasets
The proposed GCP method was evaluated on two widely used public benchmarks:
- CrowdAI Mapping Challenge: This dataset consists of aerial RGB images with a ground sampling distance of approximately 0.2m.
- WHU-Mix (vector): This dataset is a larger collection of aerial and satellite images from various sources, offering a wider range of spatial resolutions and geographical coverage.
Key Insights and Ideas
- Dynamic Programming for Global Optimization: The paper highlights the power of dynamic programming to find a globally optimal solution for polygon simplification, moving beyond the limitations of greedy or local methods.
- Collinearity as a Key Feature: Recognizing that buildings often have straight edges, the method explicitly optimizes for preserving collinearity, leading to more natural and simplified polygonal representations.
- End-to-End Trainability: Integrating the global collinearity loss allows the entire pipeline, from mask generation to polygon simplification, to be trained jointly, leading to a more cohesive and effective system.
- General Applicability: The collinearity-aware simplification module shows promise as a standalone tool for simplifying arbitrary polylines, outperforming traditional methods.
Key Results
- State-of-the-Art Performance: GCP achieved state-of-the-art results on the CrowdAI dataset for AP, AP50, and AP75 metrics, outperforming previous methods.
- Competitive Results on WHU-Mix: On the WHU-Mix dataset, GCP demonstrated superior performance, particularly in AP values, indicating better precision at larger IoU thresholds.
- Compact Polygons: GCP consistently produced polygons with a lower N-ratio (fewer vertices than ground truth) compared to other methods, indicating more compact representations.
- Improved Accuracy: The method achieved lower Maximum Tangent Angle Error (MTA) values, suggesting better geometric similarity to ground truth contours.
- Efficiency: While not the fastest component, the GCP simplification module's computational cost was comparable to other stages of the pipeline, and it can be accelerated with parallel processing.
Limitations and Future Work
- Rounded Structures: The paper acknowledges that GCP struggles with highly rounded building structures, approximating them with piecewise linear representations that can lead to fewer vertices and less accurate outlines.
- Dependency on Mask Quality: The accuracy of the predicted polygons is inherently dependent on the quality of the initial binary masks. If masks fail to capture building geometries accurately, polygon extraction will also suffer.
- Future Work Suggestions: The authors suggest adaptively adjusting the regularization coefficient ($\lambda$) based on building type to better handle curved structures. They also propose integrating more powerful instance segmentation models to improve mask quality.
Summary updated