Global Collinearity-aware Polygonizer for Polygonal Building Mapping in Remote Sensing

TL;DR

This work tackles polygonal building mapping from remote sensing masks by introducing the Global Collinearity-aware Polygonizer (GCP), a three-stage framework that refines mask contours with a transformer-based polyline regression and then applies a globally optimal, collinearity-aware polygon simplification via dynamic programming. A differentiable global collinearity loss, $oldsymbol{ ext{L}}_{ ext{gc}}$, ties the refinement and simplification into end-to-end training, improving polygon compactness (low N-ratio) while preserving geometry. Evaluations on CrowdAI and WHU-Mix-vector benchmarks show state-of-the-art MS-COCO metrics and competitive IoU, with clearer, simpler polygons than prior methods; ablations confirm the contributions of the regression module, the loss, and the GCP simplification. The method also demonstrates that applying the collinearity-aware simplification to arbitrary polylines outperforms Douglas-Peucker, indicating broad applicability for vectorization in remote sensing and digital twin workflows. The authors provide code to reproduce the results, facilitating adoption and further research in polygonal building mapping.

Abstract

This paper addresses the challenge of mapping polygonal buildings from remote sensing images and introduces a novel algorithm, the Global Collinearity-aware Polygonizer (GCP). GCP, built upon an instance segmentation framework, processes binary masks produced by any instance segmentation model. The algorithm begins by collecting polylines sampled along the contours of the binary masks. These polylines undergo a refinement process using a transformer-based regression module to ensure they accurately fit the contours of the targeted building instances. Subsequently, a collinearity-aware polygon simplification module simplifies these refined polylines and generate the final polygon representation. This module employs dynamic programming technique to optimize an objective function that balances the simplicity and fidelity of the polygons, achieving globally optimal solutions. Furthermore, the optimized collinearity-aware objective is seamlessly integrated into network training, enhancing the cohesiveness of the entire pipeline. The effectiveness of GCP has been validated on two public benchmarks for polygonal building mapping. Further experiments reveal that applying the collinearity-aware polygon simplification module to arbitrary polylines, without prior knowledge, enhances accuracy over traditional methods such as the Douglas-Peucker algorithm. This finding underscores the broad applicability of GCP. The code for the proposed method will be made available at https://github.com/zhu-xlab.

Figures (7)

• Figure 1: Flowchart of the proposed GCP. The upper section illustrates the overall pipeline, while the bottom section provides a detailed look at the architecture of two specific modules.
• Figure 2: Illustration of the initial contour generation process.
• Figure 3: Illustration of the proposed global collinearity-aware polyline simplification method. For a given polyline to be simplified $\mathbf{p}$, The middle left part of the figure lists $4$ different possible simplification plans. The bottom left part shows $2$ invalid simplification plans according to the definition in Eq. (\ref{['eq:definition']}). The right part of the figure demonstrates how the optimal solution is achieved by combining the solutions to its sub-problems according to Eq. (\ref{['eq:dp']}). Such a property is termed "optimal substructure" in Dynamic Programming.
• Figure 4: Visualized polygonal building predictions for different comparative methods on CrowdAI dataset. Red circles indicate areas where the building structure is not accurately represented.
• Figure 5: Visualized polygonal building predictions for various ablated models on WHU-Mix (vector) dataset. Red circles indicate areas where the building structure is not accurately represented. The images in the first two rows pertain to the Test 1 set, whereas the last two rows feature images from the Test 2 set..