SpaGBOL: Spatial-Graph-Based Orientated Localisation
This addresses the challenge of geo-localization for urban mapping and navigation by introducing a novel graph-structured dataset and method, though it is incremental in applying GNNs to this domain.
The paper tackled the problem of cross-view geo-localization in urban regions by proposing a graph-based approach, achieving state-of-the-art accuracies with relative Top-1 retrieval improvements of 11% and 50% when using bearing vector filtering on their dataset.
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.