AirLock+

IEEE Winter Conference on Applications of Computer Vision (WACV 2026) — Tucson, AZ

AirLock+: Scaling UAV-to-Satellite Image Registration for
Target Geolocalization and Geospatial Augmented Reality

Zhiyun Deng1 Austin Case1 Luis Sentis1,2
1Human Centered Robotics Laboratory, The University of Texas at Austin 2AIVE AI Systems
📄 Paper: CVF Open Access 📄 Paper: IEEE Xplore 🖼️ Poster 🗂️ Dataset

Overview

AirLock+ locates UAV-observed targets in world coordinates and overlays map-aligned graphics on live drone feeds.

20.23 m Mean geolocalization error
7.8 km Real-world trajectories
Zero-shot Unseen environment generalization
AirLock+ poster
Add your poster image at assets/poster.png
Click the poster to open the full-resolution PDF.

Dataset

We release the first air-to-ground target geolocalization dataset, with paired UAV and vehicle's FPV video and geo-referenced satellite maps to support UAV-to-Map registration.

↑ High-Altitude
↓ Low-Altitude

Dataset Download

All 7 flight sequences captured in Austin, TX with a DJI Air 2S, hosted on Hugging Face.

7 Flight sequences
~53 min Total flight time
21.5 km Total flight distance
118–452 m Altitude range (AGL)
57 km/h Max vehicle speed
🛰️

Reference Maps

Geo-referenced GeoTIFF tiles at zoom 17 — satellite imagery, OpenStreetMap, satellite+streets, and streets-only layers covering all flight areas.

  • 4 map styles · GeoTIFF / TIF
  • WGS84 · Zoom level 17
Download (Hugging Face)
🚁

UAV Aerial Video

7 sequences of aerial footage from a DJI Air 2S at altitudes ranging 118–452 m AGL, with embedded SRT telemetry and extracted GPS CSV. Available as camera-native 2.7K masters and 1080p proxies, each with matching OpenCV camera calibration.

  • 2688×1512 HEVC · DJI_1001–1006
  • 1920×1080 H.264 · all 7 sequences
  • Camera calibration · YAML + NPZ + TXT
  • ~53 min · 21.5 km total
Download (Hugging Face)
🚗

Vehicle FPV Video

First-person view footage from a GoPro mounted on the target vehicle, covering 5 of the 7 flight sequences. Includes raw MP4 and GPS logs from the Phyphox app.

  • 5 sequences · MP4
  • Up to 57 km/h vehicle speed
Download (Hugging Face)
📍

Time-Synchronized GPS

Per-frame CSV files aligning drone telemetry to the UAV video timeline, with vehicle GPS linearly interpolated onto the same axis. Includes a manifest and generation script.

  • 7 sequences · CSV
  • Drone + vehicle · ~60 Hz
Download (Hugging Face)
🗺️

Position-Map Videos

UAV-centered map videos rendered over the satellite+streets basemap, showing a 350 × 350 m window that follows the UAV. Cyan marks the UAV, red marks the vehicle, and each clip matches the duration of its source UAV video to within one frame.

  • 7 sequences · MP4 (H.264)
  • 720×720 · 30 fps · zoom 17 basemap
Download (Hugging Face)

Or pull the whole dataset from the command line: hf download zhiyundeng/AirLock-WACV2026 --repo-type dataset --local-dir AirLock

Please don’t hesitate to reach out to Zhiyun Deng at zdeng@utexas.edu if you have any questions about the dataset.

BibTeX

@inproceedings{deng2026airlock+,
  title={AirLock+: Scaling UAV-to-Satellite Image Registration for Target Geolocalization and Geospatial Augmented Reality},
  author={Deng, Zhiyun and Case, Austin and Sentis, Luis},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={3340--3349},
  year={2026}
}