SpaGBOL: Spatial-Graph-Based Orientated Localisation

TL;DR

SpaGBOL introduces a graph-based approach with a novel dataset and GNN architecture for cross-view geo-localisation, significantly improving accuracy by leveraging spatial relationships and neighbourhood filtering.

Contribution

It presents the first graph-structured dataset, applies GNNs to exploit spatial correlations, and introduces a new retrieval filtering method based on neighbourhood bearings.

Findings

Achieves 11% top-1 accuracy improvement over previous methods.

Demonstrates 50% accuracy boost using bearing vector filtering.

Introduces a novel graph dataset for cross-view geo-localisation.

Abstract

Cross-View Geo-Localisation within urban regions is challenging in part due to the lack of geo-spatial structuring within current datasets and techniques. We propose utilising graph representations to model sequences of local observations and the connectivity of the target location. Modelling as a graph enables generating previously unseen sequences by sampling with new parameter configurations. To leverage this newly available information, we propose a GNN-based architecture, producing spatially strong embeddings and improving discriminability over isolated image embeddings. We outline SpaGBOL, introducing three novel contributions. 1) The first graph-structured dataset for Cross-View Geo-Localisation, containing multiple streetview images per node to improve generalisation. 2) Introducing GNNs to the problem, we develop the first system that exploits the correlation between node proximity and feature similarity. 3) Leveraging the unique properties of the graph representation - we demonstrate a novel retrieval filtering approach based on neighbourhood bearings. SpaGBOL achieves state-of-the-art accuracies on the unseen test graph - with relative Top-1 retrieval improvements on previous techniques of 11%, and 50% when filtering with Bearing Vector Matching on the SpaGBOL dataset.