Contrastive learning for one-shot building shape recognition using vector polygon transformers

Accurate building shape recognition is essential for cartographic generalization, urban planning, and geographic analysis, but existing methods struggle with numerous categories and limited samples. This paper presents a novel contrastive learning-based method for building shape recognition that imp...

Full description

Saved in:
Bibliographic Details
Main Authors: Longfei Cui, Haizhong Qian, Junkui Xu, Chao Li, Xinyu Niu
Format: Article
Language:English
Published: Taylor & Francis Group 2025-12-01
Series:Geocarto International
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/10106049.2025.2471087
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate building shape recognition is essential for cartographic generalization, urban planning, and geographic analysis, but existing methods struggle with numerous categories and limited samples. This paper presents a novel contrastive learning-based method for building shape recognition that improves accuracy and broadens recognizable types in one-shot scenarios. It employs a vector polygon transformer deep neural network (VPT-DNN) for automatic feature extraction, avoiding manual calculation. By employing contrastive learning, the model effectively distinguishes between various building shapes in an unsupervised manner, requiring only a single labeled sample per shape category. Experiments demonstrate 85.1% average accuracy on a 10-category dataset, surpassing existing few-shot methods. The model generalizes effectively, achieving 87.9% accuracy on an unseen European dataset without retraining. The adoption of this methodology reduces costs associated with manual operations and enhances the overall process of building recognition and classification.
ISSN:1010-6049
1752-0762