Chapter 2 - WACV 2025 Paper Accepted! 🎉

🌍🚗 SpaGBOL: Spatial-Graph-Based Orientated Localisation 📡🗺️

Tavis Shore Oscar Mendez Simon Hadfield

Centre for Vision, Speech, and Signal Processing (CVSSP)

University of Surrey, Guildford, GU2 7XH, United Kingdom

📓 Description

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 release 🍝 SpaGBOL 🍝, the first graph-based CVGL dataset, consisting of 10 city centre graph networks across the globe. This densely sampled structured dataset will progress the CVGL field towards real-world viability.

💾 SpaGBOL: Graph-Based CVGL Dataset

SpaGBOL contains 98,855 panoramic streetview images across different seasons, and 19,771 corresponding satellite images from 10 mostly densely populated international cities. This translates to 5 panoramic images and one satellite image per graph node. Downloading instructions below.

The map below shows the cities contained in SpaGBOL v1, with the breadth and density being increased in subsequent versions.

📍 City Locations 🇬🇧🇧🇪🇺🇸🇭🇰🇸🇬🇯🇵

🧬 City Graph Representations

Here are a few of the city centre graph networks, where nodes represent road junctions, and edges represent the roads between junctions.

City of London	Manhattan Centre	Tokyo Centre

📸 Image Pair Examples

At each graph node, streetview and satellite images are collected at a ratio of 5:1 to improve training generalisation, here are some examples from across the globe.

🚶 Exhaustive / Random Depth-First Walk Generation

	Graph networks can be traversed using Breadth-First Search (BFS) or Depth-First Search (DFS). BFS explores level by level, visiting all neighbors of a node before moving deeper, using a queue. DFS dives into a branch fully before backtracking, often using a stack or recursion. BFS is ideal for shortest paths, while DFS suits tasks like cycle detection or exploring all paths.
	DFS relates to a vehicle’s movement by mimicking how a vehicle explores routes sequentially. This approach is useful for navigating unmapped areas or exploring all possible routes systematically. Reference sets contain exhaustive sampling of each node, retrieving any one of these random walks is deemed correct.

---

🧰 SpaGBOL: Benchmarking

🚧 Under Construction

🐍 Environment Setup

    conda env create -f requirements.yaml

🏭 Data Download

☕ Submodule Pretraining

👟 SpaGBOL Training

🧐 SpaGBOL Evaluation

📊 SpaGBOL: Benchmark Results

FOV	360°				180°				90°
Model	Top-1	Top-5	Top-10	Top-1%	Top-1	Top-5	Top-10	Top-1%	Top-1	Top-5	Top-10	Top-1%
CVM	2.87	12.96	21.51	28.33	2.68	9.83	15.12	20.23	1.02	5.87	10.15	14.81
CVFT	4.02	13.02	20.29	27.19	2.49	8.74	14.61	19.91	1.21	5.74	10.02	13.53
DSM	5.82	10.21	14.13	18.62	3.33	9.74	14.66		1.59	5.87	10.11	16.24
L2LTR	11.23	31.27	42.50	49.52	5.94	18.32	28.53	35.23	6.13	18.70	27.95	34.08
GeoDTR+	17.49	40.27	52.01	59.41	9.06	25.46	35.67	43.33	5.55	17.04	24.31	31.78
SAIG-D	25.65	51.44	62.29	68.22	15.12	35.55	45.63	53.10	7.40	21.76	31.14	37.14
Sample4Geo	50.80	74.22	79.96	82.32	37.52	64.52	71.92	76.39	6.51	20.61	30.31	36.12
SpaGBOL	56.48	77.47	83.85	87.24	40.88	63.79	72.88	78.28	18.63	43.20	54.05	61.20
SpaGBOL+B	64.01	86.54	92.09	94.64	52.01	82.20	89.47	93.62	-	-	-
SpaGBOL+YB	76.13	95.21	97.96	98.98	66.82	92.69	96.38	97.30	-	-	-	-

✒️ Citation

If you find SpaGBOL useful for your work please cite:

@InProceedings{Shore_2025_WACV,
    author    = {Shore, Tavis and Mendez, Oscar and Hadfield, Simon},
    title     = {SpaGBOL: Spatial-Graph-Based Orientated Localisation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025}
}

📗 Related Works

🦜 BEV-CV: Birds-Eye-View Transform for Cross-View Geo-Localisation

     

⭐ Star History