egenioussBench :
A New Dataset for Geospatial Visual Localisation

Phillipp Fanta-Jende · Francesco Vultaggio · Alexander Kern · Yasmin Loeper · Markus Gerke

Paper | Project Page | Dataset

egenioussBench couples a city-scale aerial 3D mesh, a CityGML LoD2 model, and centimetre-accurate smartphone ground-truth poses to benchmark mesh- and object-based localisation under realistic, city-scale conditions.

---

Table of Contents

• Overview
• Dataset Structure
  • Components
    • 1. Airborne 3D Mesh
    • 2. CityGML-LoD2 Model
    • 3. Smartphone Query Dataset
• Evaluation
  • Metrics
  • Running the Evaluation Script
• Submissions
• Baseline
• Citation
• Acknowledgements

---

Overview

egenioussBench is a benchmark designed to evaluate visual localisation algorithms that rely on geospatial reference data. The dataset provides:

1. A high-resolution aerial 3D mesh reconstructed from oblique imagery
2. A CityGML LoD2 building model
3. Smartphone query images captured from a Pixel8 with a tightly-coupled INS, providing cm-accurate, map-independent ground truth
4. Smartphone pose priors coming from the internal GNSS receiver

The goal is to support research on scalable localisation pipelines that operate at city scale and across different reference representations.

---

Dataset Structure

egenioussBench/
├── lod2/                       # CityGML LoD2 model of Braunschweig
├── mesh/                       # Airborne 3D mesh (7.5 cm GSD)
├── metadata
│   └── camera_parameters.txt
└── query_images
    ├── test/                   # Test images
    ├── test_android_poses.csv
    ├── val/                    # Validation
    ├── val_android_poses.txt
    └── val_gt_poses.txt

Split	Purpose	Size	GT Available?
Validation	method development	412 seq. images	✓
Test	leaderboard evaluation	42 non-co-visible images	✗

The test split is explicitly non-co-visible to enforce cold-start localisation. The dataset is available on Zenodo

Components

1. Airborne 3D Mesh

• Derived from oblique imagery (UltraCam Osprey 4.1)
• ≈1550 m AGL
• 7.5 cm GSD (nadir)
• Georeferencing accuracy ≈1 GSD (XY) / 1.5 GSD (Z)

Provides a realistic, deployable reference model for cross-view localisation.

2. CityGML LoD2 Model

• Official city model of Braunschweig
• Footprints from cadastral data
• Generalised roof shapes
• Typical corner accuracy ≈10 cm relative to mesh

Represents textureless, low-detail geometry for object-based localisation.

3. Smartphone Query Dataset

• 2709 RGB images collected in January 2024

• Resampled to 960×1280 px (~4 cm GSD)

• PPK + GCP/CP-aided bundle adjustment

• Final pose accuracy:

  • 4 cm (XY) / 7 cm (Z) mean
  • 0.04° mean orientation error

---

Evaluation

The benchmark evaluates 6-DoF camera poses predicted for each query image.

Participants submit a CSV file containing:

# Ground truth poses for validation images
# IMAGE_NAME qw qx qy qz UTMx UTMy alt

We use the same camera coordinate definition as Pix4D, for reference please look at the documentation.

Metrics

We report:

• Binned recall at:

  • 0.5 m / 2°
  • 2 m / 5°
  • 5 m / 10°

• Outliers

• Median translation error

• Median rotation error

• RMSE translation error

• RMSE rotation error

Mesh-based and LoD2-based methods are evaluated separately.

Running the evaluation script

We provide a lightweight Python evaluation script to self validate on the validation script, the same code will be used to evaluate the test split:

python eval.py \
    --pred poses.csv \
    --gt val/poses_gt.csv \
    --config eval_config.yaml \
    --visualize true
    --experiment debug

---

Submissions

Submissions should be sent by email to egeniouss@ait.ac.at.
Evaluation results will be returned via the same address. Multiple submissions are allowed, but only one submission per day will be evaluated per team.

By submitting, participants grant the organizers permission to publish the resulting scores on the public leaderboard.

Baseline

We include simple reference baselines demonstrating usage of the dataset. These currently include:

• Mesh-based baseline Based on Meshloc, explains how to

---

Citation

If you use egenioussBench in research, please consider citing:

@article{fanta-jende2025egenioussBench,
  title={egenioussBench: A New Dataset for Geospatial Visual Localisation},
  author={Fanta-Jende, Phillipp and Vultaggio, Francesco and Kern, Alexander and Loeper, Yasmin and Gerke, Markus},
  year={2025}
}

---

Acknowledgements

This work is part of the EU-Horizon egeniouss project.