Shore Tavis ; Mendez Oscar ; Hadfield Simon

Abstract

Cross-View Geo-Localisation within urban regions is challenging in part dueto the lack of geo-spatial structuring within current datasets and techniques.We propose utilising graph representations to model sequences of localobservations and the connectivity of the target location. Modelling as a graphenables generating previously unseen sequences by sampling with new parameterconfigurations. To leverage this newly available information, we propose aGNN-based architecture, producing spatially strong embeddings and improvingdiscriminability over isolated image embeddings. We outline SpaGBOL,introducing three novel contributions. 1) The first graph-structured datasetfor Cross-View Geo-Localisation, containing multiple streetview images per nodeto improve generalisation. 2) Introducing GNNs to the problem, we develop thefirst system that exploits the correlation between node proximity and featuresimilarity. 3) Leveraging the unique properties of the graph representation -we demonstrate a novel retrieval filtering approach based on neighbourhoodbearings. SpaGBOL achieves state-of-the-art accuracies on the unseen test graph- with relative Top-1 retrieval improvements on previous techniques of 11%, and50% when filtering with Bearing Vector Matching on the SpaGBOL dataset.