Contrastive Graph Autoencoder for Geometric Polygon Retrieval from Building Datasets

Abstract

Retrieval of polygon geometries with similar shapes from maps is a challenging geographic information task. Existing approaches can not process geometry polygons with complex shapes, (multiple) holes and are sensitive to geometric transformations (e.g., rotation and reflection). We propose Contrastive Graph Autoencoder (CGAE), a robust and effective graph representation autoencoder for extracting polygon geometries of similar shapes from real-world building maps based on template queries. By leveraging graph message-passing layer, graph feature augmentation and contrastive learning, the proposed CGAE reconstructs graph features of w.r.t input graph representations of polygons. The CGAE outperforms existing graph-based autoencoder on multiple polygon datasets in the task of similar shape retrieval of polygons. Experimentally, we show that our approach is capable of identifying and extracting similar complex polygon geometries with or without holes from polygonal datasets, based on template queries. Further experiments demonstrate that our approach is highly robust to geometrical transformations in contrast to existing GAE model, indicating the strong generalizability and versatility of CGAE in identifying complex real-world building footprints.