Request the dataset
USCORE

Commercial licence · for computer vision teams

uscore_uav_28x10k

10,000 curated frames from onboard UAV cameras: 28 drone classes, thermal and daylight, labels checked twice.

  • 10 000frames
  • 28UAV classes
  • 312source flights
  • YOLO · COCOlabel formats

€800perpetual licence for one organisation, excl. VAT

Request a licence

01 / Classes

28 classes, mapped to an enemy UAV reference catalog

Reconnaissance and multi-role UAVs — the ones intercepted in the air. Class names are in Latin script, like the labels in the background video.

Reconnaissance · 24

  1. SUPERCAM_S350650
  2. ZALA_Z_16530
  3. ORLAN_30490
  4. MERLIN_VR450
  5. KNYAZ_VESHCHIY_OLEG430
  6. SUPERCAM_S180380
  7. IZDELIE_56_ZALA_KUB360
  8. ZALA_Z_16_2330
  9. ELERON_3310
  10. AYAKS_V300290
  11. GRANAT_4280
  12. SKAT_350M260
  13. ZALA_Z_20250
  14. ALBATROS_M5230
  15. SARYCH_2220
  16. ELERON_10210
  17. SHTORM200
  18. FENIKS190
  19. DRAGON_F320180
  20. SH_10175
  21. ELERON_7170
  22. KARTOGRAF165
  23. ELERON_16_VALDAY160
  24. SKAT_450M160

Multi-role · 4

  1. ORLAN_10720
  2. MOLNIYA_2420
  3. GERBERA350
  4. DELTA240
  1. —decoy frames with no drones1 200

Frames per class. The largest class is 4.5× the smallest; even the rarest types have at least 160 frames.

02 / Quality

The metrics we would ask of anyone else’s dataset

Four angles on quality — and how every number was measured.

Annotation

  • 0.92mean IoU between two independent annotators1,000 frames annotated twice
  • 0.4%missed targetssecond pass over a control sample
  • 0.96class agreement, Cohen’s kappaannotator vs. an expert with the catalog

Curation

  • 1:40one frame kept for every forty raw≈400k raw frames in
  • 0near-duplicatespairs with embedding cosine similarity above 0.95 removed
  • 0flights shared by train and testsplit by flight, not by frame — no leakage

Diversity

  • 312source flightsdifferent days, altitudes and airframes
  • 31%night, fog, rain or low contrastfrom per-frame metadata
  • 2.4 kmfarthest targetmedian 610 m, from onboard telemetry

Difficulty

  • 61%small targetsunder 32×32 px, COCO definition
  • 12%decoy frames with no dronesbirds, glare, clouds, helicopters
  • 0.48mAP@50–95 of a YOLOv8s baselineon test — plenty of headroom for your model

Sensor

  • thermal 46%
  • RGB 54%

Conditions

  • day 69%
  • night 19%
  • fog, rain 12%

Target size, px

  • <16 22%
  • 16–32 39%
  • 32–96 29%
  • >96 10%

03 / Format

Unpack it and start training

We collected this dataset for our own detection and tracking systems. These are not stock images but frames from real flights — with noise, glare and small targets at the edge of visibility.

  • Boxes and classes in YOLO and COCO JSON

  • Train / val / test split 70 / 15 / 15 by flight

  • Per-frame metadata: sensor, altitude, time of day, conditions

  • Class descriptions linked to the catalog

  • Decoy frames marked separately

  • Methodology for every quality metric

04 / What for

From the first experiment to a field pilot

  1. A baseline in one evening

    Train a detector on our frames and see how it holds a target in thermal imagery.

  2. Check your own model

    Run your model on the held-out split and see where it loses objects.

  3. Fine-tune for hard conditions

    Add our frames to yours so the model stops confusing birds with drones — and an Orlan with a Molniya.

05 / Licence

One price for the whole dataset — no subscription, no per-frame fees

Two licence options. Payment by invoice, access after payment; prices exclude VAT.

  • Commercial

    €800

    One organisation, perpetual. Train and ship your own models without limits; the frames themselves may not be passed on to third parties.

  • Extended

    on request

    Several legal entities, additional classes, new frames as they arrive, or bundling into your product. Discussed individually.

06 / Request

Licence requests go through your work email

We license the dataset to companies and research teams, so we need an address on your organisation’s domain. We reply with the licence agreement and an invoice within one business day.

We use the address only to reply about the licence and for occasional news about USCORE datasets.