A Connected Building Landscape dataset for Instance Segmentation.

Area of Science:

• Computer Vision
• Urban Planning
• Architectural Analysis

Background:

• High-density East Asian cities possess unique architectural characteristics.
• A scarcity of high-quality instance segmentation data hinders research in these urban environments.
• Existing datasets do not adequately capture the complexity of East Asian urban landscapes.

Purpose of the Study:

• Introduce the Connected Building Landscape dataset, a novel resource for East Asian urban environments.
• Facilitate training and benchmarking of segmentation models for architectural morphology analysis.
• Provide tools for urban planners to generate hazard distribution maps and cultural heritage preservation reports.

Main Methods:

• The dataset comprises 2,801 JPEG images (1024 x 768 pixels) with polygonal segmentation masks.
• Annotations are provided in a standard JSON format for seamless model integration.
• Baseline experiments were conducted using Mask R-CNN and Mask2Former models.

Main Results:

• The Connected Building Landscape dataset captures diverse architectural styles and complex spatial layouts.
• Baseline experiments demonstrated the dataset's quality and robustness for segmentation tasks.
• The dataset enables a deeper understanding of East Asian architectural morphology.

Conclusions:

• The Connected Building Landscape dataset is a valuable resource for computer vision and urban planning research.
• This dataset supports more precise urban planning tools and detailed analytical reports.
• It advances the study of East Asian urban environments and architectural characteristics.