arXiv:2610.40455 WACV 2027 CC BY-NC 4.0

KITScenes Aerial

A Unified Benchmark for Aerial-Enhanced Autonomous Driving

20 cm/px
RGB orthophotos
8 cm/px
learned features
3 Tasks
unified ground-aerial benchmark

KITScenes Aerial extends KITScenes Multimodal with spatially aligned aerial imagery and learned feature representations, complementing its synchronized onboard sensor data, vehicle poses, and rich Lanelet2 HD map annotations. By exposing road structure beyond the immediate perception range of onboard sensors, it provides a unified foundation for aerial-enhanced autonomous driving research.

The extension combines 20 cm/px RGB orthophotos with neural representations derived from 8 cm/px aerial imagery. It supports a wide range of aerial-enhanced automated driving research, including aerial-only map construction, aerial-assisted online HD mapping, and cross-view localization.

Data access: The Aerial extension is available through the existing KITScenes API and dataset distribution.

Part of Multimodal v2.0.0

The DOP20 orthophotos come with every download; the DINOv3 feature tiles with the full download (kitscenes download $KITSCENES_ROOT) or with --aerial-features.

Download on HuggingFace Read the Paper arXiv ↗

Dataset Samples

KITScenes Aerial sample — camera views with LiDAR and HD map overlays, HD map on 20 cm/px aerial imagery, 8 cm/px imagery and learned features
KITScenes Aerial sample — urban intersection with ground-level sensor data, HD map, and aligned aerial representations
KITScenes Aerial sample — curved road through a park with ground-level sensor data, HD map, and aligned aerial representations

20 cm/px RGB orthophotos

Freely available DOP20 orthophotos on a north-up tile grid.

DINOv3 feature tiles

Spatially continuous features from 8 cm/px municipal orthophotos.

Upcoming

UNet++ feature crops

Ego-centered 60 × 30 m crops aligned with the onboard frame.

An upcoming addition, not yet part of v2.0.0.

Crop extraction tools

Per-frame, ego-aligned crops with configurable extent up to 240 × 60 m.

Citation

If you use KITScenes Aerial in your research, please cite our works:

The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset

arXiv:2606.02956

@misc{schwarzkopf2026roadaheadautonomousdriving, title = {The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset}, author = {Richard Schwarzkopf and Fabian Immel and Alexander Blumberg and Jonas Merkert and Nils Rack and Kaiwen Wang and Fabian Konstantinidis and Julian Truetsch and Carlos Fernandez and Annika Bätz and Kevin Rösch and Marlon Steiner and Willi Poh and Yinzhe Shen and Royden Wagner and Felix Hauser and Dominik Strutz and Jaime Villa and Gleb Stepanov and Holger Caesar and Ömer Şahin Taş and Frank Bieder and Jan-Hendrik Pauls and Christoph Stiller}, year = {2026}, eprint = {2606.02956}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2606.02956} }

A Unified Benchmark for Aerial-Enhanced Autonomous Driving: The KITScenes Aerial Dataset

arXiv:2610.40455

@misc{lengerer2026kitscenesaerial, title = {A Unified Benchmark for Aerial-Enhanced Autonomous Driving: The KITScenes Aerial Dataset}, author = {Daniel Lengerer and Richard Schwarzkopf and Jonas Merkert and Fabian Immel and Rupert Polley and Frank Bieder and Jan-Hendrik Pauls and J. Marius Zöllner and Klaus Bogenberger and Christoph Stiller and Carsten Markgraf}, year = {2026}, eprint = {2610.40455}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2610.40455} }

Team

Daniel Lengerer
Daniel Lengerer
THA

Lead Author · Research Direction

Richard Schwarzkopf
Richard Schwarzkopf
FZIKIT

Project Lead · Research Coordination

Jonas Merkert
Jonas Merkert
KIT
Fabian Immel
Fabian Immel
FZIKIT
Rupert Polley
Rupert Polley
FZI
Frank Bieder
Frank Bieder
FZIKIT
Jan-Hendrik Pauls
Jan-Hendrik Pauls
KITPD
J. Marius Zöllner
J. Marius Zöllner
FZIKIT

Supervisor

Klaus Bogenberger
Klaus Bogenberger
TUM

Supervisor

Christoph Stiller
Christoph Stiller
FZIKIT

Supervisor

Carsten Markgraf
Carsten Markgraf
THA

Supervisor

Acknowledgements

The authors gratefully acknowledge the scientific support and HPC resources provided by the Erlangen National High Performance Computing Center (NHR@FAU) of the Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) under the BayernKI project v153eb. BayernKI funding is provided by Bavarian state authorities.

The research leading to these results is partially funded by the German Federal Ministry for Economic Affairs and Energy within the project NXT GEN AI METHODS (Grant Agreement No. 19A23014), and by the Shift2SDV project (Grant Agreement No. 101194245), co-funded by the European Union through the Chips Joint Undertaking and its members. The authors gratefully acknowledge the computing time provided on the high-performance computer HoreKa by the National High-Performance Computing Center at KIT (NHR@KIT). This center is jointly supported by the Federal Ministry of Education and Research and the Ministry of Science, Research and the Arts of Baden-Württemberg, as part of the National High-Performance Computing (NHR) joint funding program. HoreKa is partly funded by the German Research Foundation (DFG). The authors also gratefully acknowledge the support of the Large Scale Data Facility (LSDF) at the Karlsruhe Institute of Technology, funded by the Ministry of Science, Research and the Arts of Baden-Württemberg and by the Federal Ministry of Education and Research.