Contrastive Graph Autoencoder for Shape-based Polygon Retrieval from Large Geometry Datasets

Zexian Huang, Kourosh Khoshelham, Martin Tomko

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., rotations). 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 layers, graph feature augmentation and contrastive learning, the proposed CGAE embeds highly discriminative latent embeddings by reconstructing graph features w.r.t. the graph representations of input polygons, outperforming existing graph-based autoencoders (GAEs) in geometry retrieval of similar polygons. Experimentally, we demonstrate this capability based on template query shapes on real-world datasets and show its high robustness to geometric transformations in contrast to existing GAEs, indicating the strong generalizability and versatility of CGAE, including on complex real-world building footprints.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

As instructed by the AE's decision, we have made the following changes to the camera-ready submission:
- Fixed typos in Abstract, Introduction and Conclusion sections.
- Corrected grammatically-incorrect sentences in sections Introduction, Discussion and Appedix at p. 15.
- Conducted a proofreading of the manuscript to align the claims made for CGAE with empirical results from experiments.
- Added a link to code for model implementation and reproducing experiment results. See Reproducibility Statement.

Code: https://github.com/zexhuang/CGAE

Assigned Action Editor: ~Nicolas_THOME2

Submission Number: 2082