EgenioussBench: Geospatial Visual Localisation

• EgenioussBench is a benchmark for geospatial visual localisation that targets estimating a full 6-DoF camera pose for ground-level images using pre-existing geospatial assets.
• It leverages an airborne 3D mesh and a CityGML LoD2 model, reflecting realistic urban mapping conditions without relying on classical SfM reconstructions.
• The benchmark enforces a cold-start, non-co-visible evaluation setup to promote scalable urban localisation research with standard mapping products.

Searching arXiv for the benchmark paper and closely related localization work to ground the article in current literature. Searching for "egenioussBench geospatial visual localisation" on arXiv. EgenioussBench is a benchmark for geospatial visual localisation that couples a city-scale airborne 3D mesh, a CityGML LoD2 3D city model, and ground-level smartphone imagery with centimetre-accurate, map-independent poses in the same urban area of Braunschweig, Germany. It is designed around real geospatial mapping assets rather than classical Structure-from-Motion reconstructions, and targets estimation of a full 6-DoF camera pose for a ground-level image using deployable reference data rather than a map built from those same images. The benchmark therefore operationalises a cross-view and cross-domain localisation setting in which airborne or city-model representations must be matched against real smartphone photographs under realistic constraints of scale, accuracy, and storage (Fanta-Jende et al., 6 May 2026).

1. Problem setting and departure from SfM

The central task in EgenioussBench is to estimate a 6-DoF camera pose for a ground-level image using only pre-existing geospatial assets, specifically meshes and 3D city models, without relying on an SfM map built from the query imagery. This differs from traditional SfM benchmarks, which typically assume a dense point cloud and reference imagery tightly coupled to the same image collection used for localisation. In those settings, the reference geometry and appearance are nearly ideal for the queries, and localisation accuracy is often correspondingly high (Fanta-Jende et al., 6 May 2026).

EgenioussBench makes three design commitments explicit. First, the reference data are real mapping products rather than reconstructions derived from the query images. Second, the benchmark enforces cold-start localisation by using a non-co-visible test set, so that each query must be localised individually against the map rather than via temporal propagation, tracking, or multi-view triangulation. Third, it targets scalability beyond SfM-based approaches by using assets that cities and mapping agencies may already maintain, namely an airborne 3D mesh and a compact LoD2 city model. This suggests a benchmark philosophy oriented toward deployable urban localisation systems rather than laboratory conditions optimised around bespoke reconstructions.

2. Reference geospatial assets and spatial frame

The first reference modality is an airborne-image-based 3D mesh built from oblique imagery captured in May 2023 with an UltraCam Osprey 4.1 sensor. The flying height is approximately 1550 m above ground, and the ground sampling distance is reported as 7.5 cm in nadir and 6.5 cm in oblique imagery at the centre. Georeferencing accuracy is approximately 1 GSD in $X,Y$ and 1.5 GSD in $Z$. The textured triangulated mesh covers roughly 570 m north-south by 700 m east-west in central Braunschweig and represents buildings, streets, vegetation, and related urban surfaces at a resolution typical of professional city-scale airborne surveys (Fanta-Jende et al., 6 May 2026).

The second reference modality is a CityGML Level of Detail 2 model provided for Braunschweig by the Landesamt für Geoinformation und Landesvermessung Niedersachsen and the City of Braunschweig. Its geometry consists of building footprints with generalised prismatic roofs, derived from cadastral building outlines (ALKIS), a 5 m Digital Terrain Model, and 3D measurement data from LiDAR or image-matching point clouds. The horizontal reference is inherited from cadastre, while the vertical reference is derived from terrain heights. Empirical alignment to the mesh is reported as sampled building corners within approximately 10 cm horizontally, with height accuracy of approximately 1 m. From a localisation standpoint, the LoD2 model is a textureless, low-detail geometric representation, making image-to-model matching substantially harder than image-to-textured-mesh or image-to-SfM matching.

All data are expressed in a common geospatial reference frame: UTM zone 32N. The approximate bounding box is given by a north-west corner at $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$ and a south-east corner at $$[N: 5{,}792{,}280\ \text{m}, E: 604{,}200\ \text{m}]$$. This common frame allows absolute-world-coordinate evaluation and supports methods that incorporate geospatial priors so long as the final prediction is a 6-DoF pose in that coordinate system.

3. Query imagery and map-independent ground truth

The query dataset comprises 2,709 RGB images collected in January 2024 in a dense urban area of central Braunschweig. The sensor is a handheld smartphone rigidly mounted to a tactical-grade INS, and the images are resampled to $960\times1280$ pixels. The approximate ground sampling distance at typical ranges is about 4 cm. Although the acquisition platform included GNSS/INS, the benchmark itself is image-based: the inertial and satellite measurements are used only to generate accurate ground truth rather than as inputs to localisation methods (Fanta-Jende et al., 6 May 2026).

Ground-truth pose generation proceeds in three stages. A precise 3D trajectory is first estimated in Post-Processed Kinematic mode by combining INS data with GNSS base and rover data. The smartphone sequence is then processed with SfM and bundle adjustment using Ground Control Points measured with RTK GNSS and Check Points measured with RTK GNSS for validation. The reported accuracy statistics are: GCP RMSE $(4, 4, 7)\,\text{cm}$ in $X,Y,Z$; CP mean error $(10, 10, 8)\,\text{cm}$ in $X,Y,Z$; image pose mean error of 8 mm in $X,Y$ and 3 mm in $Z$0; image pose standard deviation of 7 mm in $Z$1 and 1 mm in $Z$2; and orientation mean error of $Z$3 with $Z$4.

A defining property of these poses is their independence from the airborne mesh and the LoD2 model. The benchmark paper explicitly frames this as avoiding the pseudo-ground-truth situation in which query images are aligned to the same model used for localisation. In effect, the reference map and the pose supervision are decoupled, so that performance is not inflated by circular alignment.

4. Non-co-visibility construction and benchmark splits

A major methodological component of EgenioussBench is the derivation of a non-co-visible query subset from the full set of 2,709 images. Using the ground-truth poses and the airborne mesh, the authors render synthetic views and associated depth maps for all viewpoints, then compute pairwise co-visibility by comparing the rendered depth maps to quantify how much of the 3D scene is observed by both images. The resulting co-visibility matrix $Z$5 is a $Z$6 matrix in which $Z$7 encodes the fraction of co-visible pixels between image $Z$8 and image $Z$9 (Fanta-Jende et al., 6 May 2026).

The benchmark defines image pairs with fewer than 10% co-visible pixels as non-co-visible. This is converted into an undirected graph $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$0, where each node corresponds to an image and an edge connects images that are co-visible above the threshold. The selection objective is the largest mutually non-co-visible subset, formulated as a Maximum Independent Set problem,

$$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$1

Because the problem is NP-hard, the construction uses a state-of-the-art kernelization-based solver by Hespe et al., after which the result is manually refined to remove artefacts caused by mesh errors in poorly reconstructed areas. This procedure is intended to prevent trivialisation through frame-to-frame overlap and to force single-image, cold-start localisation.

The released benchmark includes two principal splits. The test split contains 42 non-co-visible images with withheld ground truth and is used for leaderboard evaluation. The validation split contains 412 sequential images with full poses and is intended for training and tuning of pose regressors, self-validation or confidence estimation, cross-view matching studies, and related analyses. The benchmark description emphasises that the official leaderboard focuses on single-image localisation even though the validation sequence can support sequential experimentation.

5. Evaluation protocol and baseline performance

EgenioussBench is organised around a public leaderboard to which participants submit predicted 6-DoF poses for the 42 test images. Submissions are provided as a CSV containing a position vector $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$2 in UTM32N together with an orientation representation convertible to a 3D rotation. The benchmark evaluates performance using binning metrics at multiple pose-error thresholds and global statistics including median error, RMSE, and outlier ratio (Fanta-Jende et al., 6 May 2026).

The stated pose-accuracy bins are the fraction of test images satisfying three threshold pairs: $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$3, $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$4, and $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$5. The benchmark also reports overall median translation and rotation errors. The abstract additionally states that global statistics include RMSE and outlier ratio. For fairness, mesh-based and LoD2-based localisation are evaluated separately while using identical thresholds and metrics, since the two reference modalities differ substantially in geometric richness and visual informativeness.

The benchmark paper reports baseline results for two mesh-based methods: MeshLoc (Panek et al., 2022) and the authors’ previous geospatial mesh-based localisation method (Vultaggio et al., 2024). Because rendering from the original airborne acquisition poses at 1,550 m altitude is unsuitable for ground localisation, both baselines use the same initialisation strategy: view sampling along the road network extracted from OpenStreetMap, rendering synthetic views from sampled poses on the mesh, retrieval using CosPlace global descriptors fine-tuned for synthetic images as in MeshVPR to select the 500 most visually similar renderings, further filtering to 50 candidates near the smartphone’s GNSS pose, local feature matching with SuperPoint, and pose estimation via 2D–3D correspondences and PnP plus RANSAC. The authors’ method further uses MAGSAC++ and PnP refinements.

On the 42 non-co-visible test images, MeshLoc achieves $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$6 for the three bins, median errors of $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$7 and $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$8, and a runtime of 290 s. The authors’ method achieves $$[N: 5{,}792{,}850\ \text{m}, E: 603{,}500\ \text{m}]$$9, median errors of $$[N: 5{,}792{,}280\ \text{m}, E: 604{,}200\ \text{m}]$$0 and $$[N: 5{,}792{,}280\ \text{m}, E: 604{,}200\ \text{m}]$$1, and a runtime of 41 s. Runtime is reported but is not part of the official ranking. The benchmark paper notes that performance drops substantially relative to MeshLoc’s performance on classic datasets such as Aachen, which it interprets as evidence of the difficulty of airborne-to-ground, mesh-based localisation under realistic geospatial conditions.

6. Challenge profile, relevance, and limitations

The benchmark is explicitly designed to expose cross-view and cross-domain difficulty. Cross-view refers to the disparity between aerial acquisition viewpoints, including nadir and oblique views from high altitude, and ground-level smartphone viewpoints. Cross-domain refers to the discrepancy between rendered textured meshes or textureless LoD2 geometry and real RGB photography, compounded by different illumination, season, and sensor characteristics. These factors are amplified by the fact that the airborne mesh was not optimised for ground-level localisation, so certain ground-visible surfaces may be missing, occluded, or poorly reconstructed (Fanta-Jende et al., 6 May 2026).

The benchmark paper further notes that even renderings generated from the ground-truth pose may yield substantial pose errors because of reconstruction imperfections such as obstructions, blind spots, or inaccurate correspondences. It therefore challenges the common assumption in mesh-based localisation that the image similarity objective is minimised at the true pose. A plausible implication is that viewpoint selection itself becomes a critical subproblem: the most matchable rendered view need not coincide with the actual camera pose. For LoD2 localisation, the difficulty is sharper still, because the reference is limited to simplified, textureless building geometry and thus favours edge-, wireframe-, silhouette-, or domain-adapted matching strategies.

In practical terms, the benchmark is framed as a response to scalability constraints in city-scale localisation. It does not require a dense, feature-annotated SfM map built from street-level imagery; instead, it assumes the availability of city-scale meshes and LoD2 models, which the benchmark description characterises as storage efficient, easier to maintain, and standardised. This makes the dataset relevant to robotics, autonomous systems, augmented reality, and geospatial AI workflows that must operate over broad urban areas using official mapping products or digital-twin infrastructure.

The benchmark’s limitations are also stated directly. Geographic coverage is confined to a single city area of roughly $$[N: 5{,}792{,}280\ \text{m}, E: 604{,}200\ \text{m}]$$2. The current query data come from a single smartphone acquisition campaign in January 2024, with limited weather and time-of-day diversity. Only LoD2 building models are included; richer LoD levels and additional object classes such as trees or street furniture are absent. Mesh imperfections are retained rather than corrected, because they are intended to reflect real-world airborne mesh quality. Planned or suggested future directions include extending the geographic scope within Braunschweig and to additional cities, adding other query modalities such as UAVs, bicycles, multi-camera rigs, or panoramic cameras, incorporating richer LoD levels and additional geospatial assets, and organising workshops or special issues to consolidate evaluation practice.

Access is provided through the project website and a public GitHub repository containing the data and evaluation code. The released resources include the airborne mesh, the CityGML LoD2 model, the query images, and the validation split poses, together with scripts and documentation for data parsing, baseline reproduction, and synthetic-view generation. In aggregate, EgenioussBench defines a rigorous benchmark for large-scale visual localisation grounded in deployable geospatial assets and explicitly structured to measure performance under non-co-visible, cold-start conditions rather than SfM-favourable overlap (Fanta-Jende et al., 6 May 2026).