mehmetkeremturkcan/UrbanAnonymizer-Dataset

UrbanAnonymizer Dataset

Mehmet Kerem Turkcan
Columbia University

The UrbanAnonymizer Dataset contains 134,444 images with 394,800 face and license plate boxes in COCO format, compiled from Open Images, WIDER FACE, DARK FACE, CRPD and BirdsEye-RU. Each image lists the classes whose annotation is complete (verified_classes), so detectors can be trained and evaluated on partially labeled sources without treating unlabeled objects as background. Verified negative images, which contain neither class, reduce false masking of object-free regions.

UrbanAnonymizer was trained on this dataset. Its D-FINE-X model reaches face AP₅₀ of 83.4 and plate AP₅₀ of 92.9 on the held-out test split.

Dataset summary

Split	Images	Face boxes	Plate boxes	Sources
train	119,384	307,093	40,334	Open Images, WIDER FACE, DARK FACE, CRPD, BirdsEye-RU, Open Images negatives
validation	2,431	8,249	154	Open Images, WIDER FACE, DARK FACE, BirdsEye-RU, Open Images negatives
test	7,760	29,738	3,660	Open Images, WIDER FACE, DARK FACE
test_crpd	2,507	–	4,364	CRPD test
test_birdseye_ru	508	1,208	–	BirdsEye-RU
test_openimages_negative	1,854	–	–	Open Images negatives

The splits share no images. train also contains 5,187 crops of Open Images, WIDER FACE and DARK FACE images around small or densely packed faces (tile_kind, parent_image_id), which keep these faces visible at the training resolution.

Training sources

Source	Train images	Classes	Content
Open Images	47,178	face, plate	Human-face and vehicle-registration-plate boxes, verified per class
Open Images negatives	24,029	face, plate (absent)	Images verified to contain no faces or no plates
CRPD	24,970	plate	Chinese road license plates from traffic cameras
WIDER FACE	14,041	face	Faces in crowded and unconstrained scenes
DARK FACE	7,233	face	Faces in low light at night
BirdsEye-RU	1,933	face	Faces seen from elevated cameras

Classes

ID	Name	Description
0	face	Human face, any pose, including partially occluded faces
1	license_plate	Vehicle registration plate

COCO format

images:
  id, file_name, width, height
  source, source_id, source_split, source_url
  verified_classes                     classes whose boxes are complete for this image
  license, attribution                 per-image license and author (Open Images)
  canonical_image_id, parent_image_id, tile_kind

annotations:
  id, image_id, category_id, bbox, area, iscrowd

categories:
  {id: 0, name: face}, {id: 1, name: license_plate}

Boxes use COCO [x, y, width, height] in pixels. A class absent from verified_classes is unlabeled for that image; training losses and evaluation should ignore it there.

Repository layout

annotations/
  train.json
  validation.json
  test.json
  test_crpd.json
  test_birdseye_ru.json
  test_openimages_negative.json
data/
  train-00000.tar ... train-00019.tar
  validation-00000.tar
  test-00000.tar
  test_crpd-00000.tar
  test_birdseye_ru-00000.tar
  test_openimages_negative-00000.tar
manifest.jsonl
README.md

Each TAR stores images at their file_name paths (images/<source>/<name>). manifest.jsonl maps every image to its split, shard and SHA-256 digest.

Loading the data

hf download mehmetkeremturkcan/UrbanAnonymizer-Dataset --repo-type dataset --local-dir urbananonymizer
cd urbananonymizer
for shard in data/*.tar; do tar -xf "$shard"; done

from pathlib import Path

from PIL import Image
from pycocotools.coco import COCO

root = Path("urbananonymizer")
coco = COCO(root / "annotations" / "train.json")
image_id = coco.getImgIds()[0]
record = coco.loadImgs([image_id])[0]
image = Image.open(root / record["file_name"]).convert("RGB")
boxes = coco.loadAnns(coco.getAnnIds(imgIds=[image_id]))
complete_classes = record["verified_classes"]

Benchmark results

Face and plate AP₅₀ of UrbanAnonymizer and public baselines, all scored by the same evaluator (full tables in the model repository):

Method	test faces	test plates	test_crpd plates
EgoBlur Gen2	65.2	89.8	94.8
EgoBlur Gen1	64.4	89.3	96.0
CenterFace (deface)	69.6	–	–
YOLO11x-plate	–	79.9	90.6
UrbanAnonymizer, D-FINE-X @1408	83.4	92.9	98.0

The external benchmarks UFDD and PP4AV are not redistributed here; obtain them from their publishers.

Licenses and citations

Images and annotations keep the terms of their original sources. Several sources restrict use to non-commercial research. Use of this compilation must comply with every source's terms.

Source	Terms recorded for this release	Primary record
Open Images	Annotations CC BY 4.0; images listed as CC BY 2.0, with per-image author and license in attribution	storage.googleapis.com/openimages
WIDER FACE	Non-commercial research use	shuoyang1213.me/WIDERFACE
DARK FACE	Research use; cite the UG2+ Prize Challenge dataset paper	DARK FACE (UG2+ Track 2)
CRPD	Research use; cite the publisher	Chinese Road Plate Dataset (CRPD)
BirdsEye-RU	Listed as CC BY-SA 4.0; publisher text also restricts use to academic research	Publisher listing

Ethical considerations

The dataset contains real faces and license plates in public images. It exists to build anonymization tools that remove this information before imagery is shared or analyzed. It must not be used for face recognition, re-identification, plate reading or surveillance.

Citation

@misc{turkcan2026urbananonymizerdataset,
  title        = {UrbanAnonymizer Dataset},
  author       = {Turkcan, Mehmet Kerem},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mehmetkeremturkcan/UrbanAnonymizer-Dataset}}
}

Users must also cite the source datasets.