[Paper Review] PolyMapper: Extracting City Maps using Polygons.

Zuoyue Li, Jan Dirk Wegner+1 more|arXiv (Cornell University)|Dec 4, 2018

Automated Road and Building Extraction39 references12 citations

TL;DR

PolyMapper proposes an end-to-end deep learning model that directly predicts vectorized city maps—building footprints and road networks—as closed polygons from aerial images, using a novel sequentialization method to unify geometric representations. It achieves state-of-the-art performance in automating large-scale topological map extraction with high fidelity to existing map services.

ABSTRACT

We propose a new approach, named PolyMapper, to circumvent the conventional pixel-wise segmentation of (aerial) images and predict objects in a vector representation directly. PolyMapper directly extracts the topological map of a city from overhead images as collections of building footprints and road networks. In order to unify the shape representation for different types of objects, we also propose a novel sequentialization method that reformulates a graph structure as closed polygons. Experiments are conducted on both existing and self-collected large-scale datasets of several cities. Our empirical results demonstrate that our end-to-end learnable model is capable of drawing polygons of building footprints and road networks that very closely approximate the structure of existing online map services, in a fully automated manner. Quantitative and qualitative comparison to the state-of-the-art also shows that our approach achieves good levels of performance. To the best of our knowledge, the automatic extraction of large-scale topological maps is a novel contribution in the remote sensing community that we believe will help develop models with more informed geometrical constraints.

Motivation & Objective

• To develop an end-to-end learnable model that bypasses pixel-wise segmentation and directly outputs vectorized city maps.
• To unify the representation of diverse urban objects (e.g., buildings, roads) using a common polygon-based format.
• To enable fully automated, large-scale extraction of topological city maps from overhead imagery.
• To improve geometric fidelity and topological consistency in remote sensing map generation.

Proposed method

• Proposes a novel sequentialization method that reformulates graph-structured urban objects into closed polygon representations.
• Designs a deep learning architecture that directly predicts polygon coordinates from aerial images, bypassing segmentation.
• Uses a differentiable polygon representation to enable end-to-end training with geometric supervision.
• Employs a unified vector-based loss function to optimize both shape accuracy and topological consistency.
• Trains and evaluates on both existing and newly collected large-scale aerial datasets of multiple cities.
• Integrates multi-scale features and spatial attention to enhance localization and boundary precision.

Experimental results

Research questions

• RQ1 Can a deep learning model directly predict vectorized urban maps (buildings and roads) as closed polygons without pixel-wise segmentation?
• RQ2 How well can a unified polygon representation preserve geometric and topological accuracy across diverse urban object types?
• RQ3 To what extent does the proposed sequentialization method improve the modeling of complex urban structures compared to traditional graph or segmentation-based approaches?
• RQ4 How does PolyMapper perform in comparison to state-of-the-art methods in terms of quantitative map quality and visual fidelity?

Key findings

• PolyMapper achieves high geometric accuracy, producing building footprints and road networks that closely approximate the structure of existing online map services.
• The model demonstrates strong generalization on both existing and self-collected large-scale datasets across multiple cities.
• Quantitative comparisons show superior performance over state-of-the-art methods in vector map extraction tasks.
• The proposed sequentialization method enables effective end-to-end training with improved shape and topology consistency.
• The approach enables fully automated, scalable generation of topological city maps from aerial imagery.
• The results confirm that vector-based modeling with polygon supervision yields more informed geometrical constraints than pixel-level segmentation.