LDPoly: Latent Diffusion for Polygonal Road Outline Extraction in Large-Scale Topographic Mapping

TL;DR

LDPoly addresses polygonal road outline extraction from high-resolution aerial imagery by introducing a dual-latent diffusion model that simultaneously denoises road masks and vertex heatmaps in a latent space conditioned on the input image $I$, followed by a polygonization step. A Channel-Embedded Fusion Module enables interaction between the two latent targets to produce geometrically regular, topologically coherent road polygons with minimal vertex redundancy, and two new metrics for polygon simplicity and boundary smoothness are proposed, including $\text{S-IoU}(m,\hat{m}) = \text{IoU}(m,\hat{m}) \cdot \text{SF}(N_{\,\hat{m}})$ and the SCR ratio. Evaluations on Map2ImLas show LDPoly outperforms seven baselines on pixel-level and geometric/topological metrics and generalizes well to unseen regions, demonstrating the promise of diffusion-based approaches for vectorized road outlines in large-scale topographic mapping. This work establishes a transferable framework for accurate, compact polygonal representations of roads, with potential impact on digital topographic databases and downstream GIS/HD-map pipelines.

Abstract

Polygonal road outline extraction from high-resolution aerial images is an important task in large-scale topographic mapping, where roads are represented as vectorized polygons, capturing essential geometric features with minimal vertex redundancy. Despite its importance, no existing method has been explicitly designed for this task. While polygonal building outline extraction has been extensively studied, the unique characteristics of roads, such as branching structures and topological connectivity, pose challenges to these methods. To address this gap, we introduce LDPoly, the first dedicated framework for extracting polygonal road outlines from high-resolution aerial images. Our method leverages a novel Dual-Latent Diffusion Model with a Channel-Embedded Fusion Module, enabling the model to simultaneously generate road masks and vertex heatmaps. A tailored polygonization method is then applied to obtain accurate vectorized road polygons with minimal vertex redundancy. We evaluate LDPoly on a new benchmark dataset, Map2ImLas, which contains detailed polygonal annotations for various topographic objects in several Dutch regions. Our experiments include both in-region and cross-region evaluations, with the latter designed to assess the model's generalization performance on unseen regions. Quantitative and qualitative results demonstrate that LDPoly outperforms state-of-the-art polygon extraction methods across various metrics, including pixel-level coverage, vertex efficiency, polygon regularity, and road connectivity. We also design two new metrics to assess polygon simplicity and boundary smoothness. Moreover, this work represents the first application of diffusion models for extracting precise vectorized object outlines without redundant vertices from remote-sensing imagery, paving the way for future advancements in this field.

Figures (20)

• Figure 1: The differences between road networks (left), road boundaries (middle), and polygonal road outlines (right).
• Figure 2: The overall architecture of our proposed method LDPoly. The double-added double-throw switch symbol indicates that during training, the corrupted latent vectors of the road mask and vertex heatmap are fed into the Channel Embedded Fusion Module (CEFM), whereas during inference, randomly generated Gaussian noise with the same size as the latent vectors is used as input to CEFM. The output road mask and vertex heatmap are further processed by a polygonization algorithm to extract the polygonal road outline.
• Figure 3: An example of the training input, including the aerial image (left), the road mask (middle), and the vertex heatmap (right), along with their corresponding latent representations and the encoder $\varepsilon$ of the autoencoder. The latent representations, with a last dimension of $c$, are visualized by mapping each channel separately to a grayscale image. In our experiments, $c=4$. Hence, each latent representation is displayed as a row of four grayscale images.
• Figure 4: The detailed structure of the Channel-Embedded Fusion Module. Here, $m'_T$, $k'_T$ and $c'$ represent the high-dimensional latent features of the road mask, vertex heatmap, and conditioning image, respectively. $\text{chn}_1$, $\text{chn}_2$, and $\text{chn}_3$ are channel-specific offsets added to their corresponding latent features.
• Figure 5: Study area map indicating the locations of Deventer, Enschede, and Giethoorn in the Netherlands. The base map imagery is accessed via Esri’s World Imagery service, which incorporates data from multiple sources including commercial satellites and national mapping agencies. Administrative boundaries are derived from GADM data https://gadm.org/.