zhiyundeng/AirLock-WACV2026

AirLock+ Dataset

Seven UAV flights over Austin, TX, each pairing nadir aerial video with geo-referenced satellite basemaps, per-frame drone telemetry, and — for five flights — first-person video and GPS from a car driving below.

Built for UAV-to-satellite image registration and air-to-ground target geolocalization: you get the aerial view, the map it should register against, and independently-measured ground truth for where the target vehicle actually was.

Accompanies the WACV 2026 paper AirLock+: Scaling UAV-to-Satellite Image Registration for Target Geolocalization and Geospatial Augmented Reality.

Project site · Paper (CVF) · IEEE Xplore

---

Poster

AirLock+ locates UAV-observed targets in world coordinates and overlays map-aligned graphics on live drone feeds. This dataset is the data release behind that system.

WACV 2026 poster — full-resolution PDF.

---

Preview

Each clip is a 30-second excerpt: the nadir UAV view fills the frame, with a moving-map inset (bottom-left) tracking the UAV and vehicle against the satellite basemap, and the vehicle's FPV camera (bottom-right) where it was recording. DJI_1003 is aerial-only — no composite edit exists for that flight.

DJI_1001 — 11:25, 445 m AGL	DJI_1002 — 6:29, 118–156 m AGL
DJI_1003 — 11:17, 441 m AGL	DJI_1004 — 7:47, 431–452 m AGL
DJI_1005 — 3:58, 149 m AGL	DJI_1006 — 8:36, 397 m AGL
DJI_1007 — 3:10, 350–413 m AGL	Full-length composites (0.4–1.3 GB each, 4.7 GB total) are in previews/full/. If the players above don't render in your browser, the clips are in previews/.

---

At a glance

Flight sequences	7
Total flight time	52 min 46 s
Total ground track	21.5 km
Altitude range (AGL)	118 – 452 m
Max vehicle speed	57 km/h
Aircraft / camera	DJI Air 2S (1″ CMOS, calibrated)
UAV video	2688×1512 HEVC + 1920×1080 H.264, 59.94 fps
Telemetry rate	~60 Hz, one sample per video frame
Capture dates	2025-08-23 (1001–1005), 2025-09-18 (1006–1007)
Location	Austin, TX — 30.258–30.302 °N, 97.784–97.722 °W
Total size	51 GB

Sequences

Seq	Duration	Alt. AGL	Track	UAV 2.7K	UAV 1080p	FPV video	Vehicle GPS	Scene
DJI_1001	11:25	445–446 m	4.48 km	✅ 7.8 GB	✅	✅	❌	Highway interchange, rail corridor
DJI_1002	6:29	118–156 m	2.65 km	✅ 4.5 GB	✅	✅	✅	Tree-lined residential, low altitude
DJI_1003	11:17	441–446 m	4.55 km	✅ 7.7 GB	✅	✅	✅	Downtown core, high-rise
DJI_1004	7:47	431–452 m	4.66 km	✅ 5.3 GB	✅	✅	✅	Waterfront, residential, parkland; fastest vehicle
DJI_1005	3:58	149–150 m	2.11 km	✅ 2.7 GB	✅	✅	✅	Commercial strip and parking
DJI_1006	8:36	397–398 m	1.71 km	✅ 5.9 GB	✅	❌	⚠️ first 482 s	University campus
DJI_1007	3:10	350–413 m	1.37 km	❌ 1080p-native	✅	❌	✅	Residential and parkland, altitude change

Read the gaps before you plan an experiment:

• DJI_1007 was recorded in the aircraft's 1080p mode, so no 2.7K master exists for it.
• DJI_1001 has vehicle video but no vehicle GPS — usable for registration, not for geolocalization error metrics.
• DJI_1006 vehicle GPS stops at ~482 s of a 517 s flight; rows past that are marked unavailable.
• DJI_1006 and DJI_1007 have no FPV video.
• 5 of 7 sequences (1002–1005, partially 1006) have both a ground-truth vehicle track and synchronized aerial video — these are the ones usable end-to-end.

Layout

previews/                     30 s clips + posters for each sequence
  full/                       full-length composite edits (6 seqs, 4.7 GB)

uav-video/
  2.7k/         DJI_100N_2.7k.mp4     2688×1512 HEVC 59.94 fps  (1001–1006)
  1080p/        DJI_100N_1080p.mp4    1920×1080 H.264 59.94 fps (1001–1007)
  calibration/  dji_air2s_2.7k.{yaml,npz,txt}, dji_air2s_1080p.yaml

vehicle-video/  DJI_100N_fpv.mp4      1920×1080 H.264, GoPro FPV (1001–1005)

telemetry/
  uav/                  DJI_100N.srt   raw DJI subtitle stream (full camera EXIF)
                        DJI_100N.csv   parsed time/lat/lon/altitude
  vehicle/              DJI_100N.csv   per-sequence vehicle GPS (Phyphox), + manifest.csv
  vehicle-raw/          <ISO-timestamp>.csv   unmodified Phyphox app exports
  vehicle-video-aligned/DJI_100N.csv   vehicle GPS re-zeroed to video start
  synchronized/         DJI_100N.csv   ★ UAV + vehicle merged on the video clock
                        manifest.csv, README.md

reference-map/
  basemap/      mapbox-satellite_z17.tif, mapbox-satellite-streets_z17.tif,
                mapbox-streets_z17.tif, openstreetmap_z17.tif, naip_z17.tif (+ .jpg)
  overlay/
    uav-path/            DJI_100N_uav-path.{png,tif}, uav-paths.geojson, index.csv
    vehicle-path/        DJI_100N_vehicle-path.{png,tif}, vehicle-paths.geojson, index.csv
    vehicle-trajectory/  DJI_100N_vehicle-trajectory.{png,tif}, …

map-video/      DJI_100N_map.mp4       720×720 moving-map render, 30 fps

The file most people want

telemetry/synchronized/DJI_100N.csv — one row per UAV video frame, with the vehicle's position interpolated onto the same clock. This is the join between aerial pixels and ground truth.

frame_index, video_time_s, timestamp, drone_lat, drone_lon, drone_altitude_m,
vehicle_lat, vehicle_lon, vehicle_altitude_m, vehicle_altitude_wgs84_m,
vehicle_speed, vehicle_data_status

vehicle_data_status is exact, interpolated, or unavailable — always filter on it before computing metrics; vehicle values are never extrapolated beyond the measured track.

import pandas as pd
df = pd.read_csv("telemetry/synchronized/DJI_1004.csv")
gt = df[df.vehicle_data_status != "unavailable"]      # rows with real ground truth
row = gt.iloc[1000]                                    # frame -> UAV pose + target position

Frame i of uav-video/*/DJI_1004_*.mp4 corresponds to frame_index == i; both video tiers share the same frame count and timebase, so the CSV applies to either.

Other formats

Path	Format	Notes
telemetry/uav/*.srt	DJI subtitle stream	Per-frame iso, shutter, fnum, ev, ct, focal_len, dzoom_ratio, latitude, longitude, altitude
telemetry/uav/*.csv	CSV	Parsed Time (s), Latitude (°), Longitude (°), Altitude (m)
telemetry/vehicle/*.csv	CSV	Phyphox GPS: timestamp, lat/lon, altitude (MSL + WGS84), speed, heading, accuracy, satellites
reference-map/basemap/*.tif	GeoTIFF	WGS84, zoom 17, covering all flight areas
reference-map/overlay/**/*.geojson	GeoJSON	Flight paths with per-feature distance, duration, sample counts
reference-map/overlay/**/index.csv	CSV	Per-sequence summary: distance, altitude min/max, start/end local time

Camera calibration

Intrinsics were estimated at the camera-native 2688×1512 and are provided as OpenCV FileStorage YAML, NumPy .npz, and plain text in uav-video/calibration/.

	2.7K master	1080p proxy
Image size	2688 × 1512	1920 × 1080
fx, fy	2470.18, 2485.77	1764.41, 1775.55
cx, cy	1316.53, 806.53	940.38, 576.09
Distortion (plumb_bob)	k1 0.088717, k2 −1.543803, p1 0.004218, p2 −0.005799, k3 5.873702	same
FOV (H × V)	57.1° × 33.8°	57.1° × 33.8°

The 1080p matrix is the 2.7K matrix scaled by 1920/2688; distortion coefficients are dimensionless and shared. Use the calibration matching the tier you downloaded.

import cv2
fs = cv2.FileStorage("uav-video/calibration/dji_air2s_2.7k.yaml", cv2.FILE_STORAGE_READ)
K    = fs.getNode("camera_matrix").mat()
dist = fs.getNode("distortion_coefficients").mat()

How the synchronization works

Drone telemetry arrives embedded in the video as an SRT stream, so it is inherently frame-aligned — no cross-device sync needed for the UAV side. The vehicle carried a phone running Phyphox logging GPS independently; its track is linearly interpolated onto the UAV video's elapsed-time axis to produce telemetry/synchronized/. manifest.csv records, per sequence, the video duration and frame count, drone sample coverage, and how many vehicle samples were exact, interpolated, or unavailable.

For DJI_1006 the vehicle track is deliberately shifted by 34.999362 s so its first sample lands at video time zero; this is why its coverage ends early.

Download

# everything (51 GB)
hf download zhiyundeng/AirLock-WACV2026 --repo-type dataset --local-dir AirLock

# just what you need
hf download zhiyundeng/AirLock-WACV2026 --repo-type dataset --local-dir AirLock \
  --include "telemetry/**" "uav-video/1080p/*" "uav-video/calibration/*"

from huggingface_hub import snapshot_download
snapshot_download(
    "zhiyundeng/AirLock-WACV2026", repo_type="dataset",
    allow_patterns=["telemetry/synchronized/*", "uav-video/1080p/*",
                    "uav-video/calibration/*", "reference-map/basemap/*"],
)

Sizes on disk: 2.7K video 34.0 GB · full previews 4.7 GB · FPV video 4.7 GB · basemaps 3.6 GB · 1080p video 3.5 GB · map videos 340 MB · map overlays 280 MB · telemetry 100 MB · preview clips 50 MB.

Start with uav-video/1080p/ + telemetry/ + reference-map/basemap/ (7.2 GB) — that's enough to reproduce registration and geolocalization experiments. Add the 2.7K masters only if you need native resolution.

Citation

@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}
}

License

CC BY 4.0 — © 2026 Zhiyun Deng, Austin Case, and Luis Sentis. Share and adapt with attribution.

Applies only to material the copyright holders can license; it grants no rights over privacy, publicity, trademarks, or patents. Basemap imagery in reference-map/basemap/ is redistributed under its providers' terms (Mapbox, OpenStreetMap contributors, USDA NAIP).

Footage was captured over public areas; incidental vehicles, buildings, and people appear. Please use it responsibly and do not attempt to identify individuals.

Contact

Zhiyun Deng — zdeng@utexas.edu