Meridian™ Blur 1
Meridian™ Blur 1 finds faces and license plates in images and video, so they can be made unrecognizable before footage is stored, shared or published. It is built and trained by Ordnary and released under the Apache 2.0 license.
At a glance
| Task | Object detection for privacy redaction |
| Classes | 0 face, 1 license_plate |
| Architecture | RF-DETR Small (Apache 2.0), fine-tuned by Ordnary |
| Input | RGB, resized to 640 x 640, plus 960 px tiles on large images |
| Formats | ONNX (fp32, int8), safetensors |
| Runs on | CPU or GPU, through ONNX Runtime |
| License | Apache 2.0 (model and code) |
Quick start
pip install onnxruntime opencv-python-headless numpy huggingface_hub
from huggingface_hub import snapshot_download
import sys, cv2
path = snapshot_download("ordnaryai/meridian-blur")
sys.path.insert(0, path)
from inference import MeridianBlur
model = MeridianBlur(path)
image = cv2.imread("street.jpg")
boxes = model.detect(image) # [(kind, x0, y0, x1, y1, score), ...]
cv2.imwrite("street-blurred.jpg", model.blur(image, boxes))
Or from the command line, for an image or a video:
python inference.py street.jpg street-blurred.jpg
python inference.py dashcam.mp4 dashcam-blurred.mp4
inference.py pads every box before blurring (35% for faces, 12% for
plates), because hair, ears and the jaw line identify people too, and on
video it keeps every box for a few frames after it was last seen, so a
face the model loses for one frame stays covered.
Scores
Measured on Open Images V7 validation and test images that were not used for training. The leak rate is the share of labeled faces and plates that end up less than 60% covered by the padded redaction. For privacy it is the number that matters: a blurred lamp post costs nothing, a visible face is the failure.
| Model | Leak, faces | Leak, plates | Leak, all | mAP50 faces | mAP50 plates |
|---|---|---|---|---|---|
| Meridian™ Blur 1 | 3.6% | 10.6% | 4.6% | 0.903 | 0.857 |
| Meridian™ Blur 1, int8 | 3.7% | 10.9% | 4.8% | 0.903 | 0.858 |
| YuNet + YOLOv9-t | 15.5% | 12.1% | 15.0% | 0.698 | 0.845 |
Leak rate by size (long side of the labeled box: small under 32 px, medium 32 to 96, large over 96):
| Model | Small faces | Medium faces | Large faces | Small plates | Medium plates | Large plates |
|---|---|---|---|---|---|---|
| Meridian™ Blur 1 | 17.6% | 4.1% | 2.1% | 44.8% | 5.3% | 1.5% |
| Meridian™ Blur 1, int8 | 18.6% | 4.3% | 2.2% | 45.8% | 5.6% | 1.5% |
| YuNet + YOLOv9-t | 26.0% | 15.0% | 14.9% | 51.3% | 6.4% | 1.4% |
Compared with other open models
Same images, same scoring code, same padding; every other model at its own default threshold.
Faces
| Model | Leak, all | Small | Medium | Large | mAP50 |
|---|---|---|---|---|---|
| Meridian™ Blur 1 | 3.6% | 17.6% | 4.1% | 2.1% | 0.903 |
| CenterFace (deface 1.5.0) | 10.8% | 16.3% | 9.6% | 10.9% | 0.784 |
| RetinaFace ResNet-50 | 11.8% | 35.5% | 13.4% | 9.0% | 0.833 |
| YuNet (OpenCV Zoo, 2023mar) | 15.5% | 26.0% | 15.0% | 14.9% | 0.698 |
| MediaPipe face detection, full range | 47.6% | 97.6% | 73.7% | 31.1% | 0.443 |
License plates
| Model | Leak, all | Small | Medium | Large | mAP50 |
|---|---|---|---|---|---|
| Meridian™ Blur 1 | 10.6% | 44.8% | 5.3% | 1.5% | 0.857 |
| YOLOv9-s 608, with 960 px tiles | 9.3% | 43.7% | 3.7% | 0.4% | 0.843 |
| YOLOv9-s 608 | 11.7% | 54.9% | 4.6% | 0.4% | 0.866 |
| YOLOv9-t 640, with 960 px tiles | 12.1% | 51.3% | 6.4% | 1.4% | 0.845 |
On faces, Meridian™ Blur 1 leaks about a third as often as the best other model tested. CenterFace leaks fewer of the smallest faces, partly because it marks many more boxes. On license plates the models are level: YOLOv9-s with tiles leaks less (9.3% against 10.6%), but more of what it blurs is not a labeled plate. MediaPipe is built for faces close to the camera. Meta's EgoBlur was not tested: its weights are only available after accepting a license form.
14,431 images (3,639 from the Open Images validation split, 10,792 from test) with 21,251 labeled faces and 3,808 labeled plates. Meridian™ Blur 1 runs at score threshold 0.3 with tiling. YuNet (OpenCV Zoo, 2023mar) and the YOLOv9-t license plate model from open-image-models run at their usual thresholds (face 0.5, plate 0.25) on the same images.
Open Images labels faces and plates in a particular style, and Meridian™ Blur 1 was trained on images labeled the same way, so this comparison favors it. A hand-checked test set from other sources is still to come.
Latency, one image on a CPU (4 cores, ONNX Runtime, median of 15 runs):
| Model | 1280 px wide | 2400 px wide |
|---|---|---|
| Meridian™ Blur 1, tiled | 895 ms | 2,077 ms |
| Meridian™ Blur 1, int8, tiled | 519 ms | 1,203 ms |
| Meridian™ Blur 1, int8, full image only | 189 ms | 194 ms |
| YuNet + YOLOv9-t | 137 ms | 330 ms |
Training data
Trained on images from Open Images V7,
classes "Human face" and "Vehicle registration plate". The annotations are
licensed CC BY 4.0 by Google; the images are licensed CC BY 2.0 by their
Flickr authors. The exact images used are listed in image_ids.txt, and
DATA_SOURCES.md has the full attribution. No images of customers were
used.
Training set: 30,158 images with 75,134 labeled faces and 7,543 labeled plates. 12,381 faces that Open Images left unlabeled were added by YuNet (OpenCV Zoo, MIT) at a high score threshold. Validation set: 800 images. Trained for 20 epochs on one NVIDIA H100 (90 minutes).
Intended use
- Blurring faces and license plates in photos, dashcam and camera footage before it is stored, shared or published.
- Privacy by design in products that handle images of the public, under the GDPR and similar laws.
Out of scope
Meridian™ Blur 1 is made to hide people, not to find them. Don't use it to:
- identify, recognize or track people, or read license plates to follow vehicles;
- build surveillance or profiling systems;
- decide anything about a person.
See Ordnary's AI Usage Policy.
Limitations
Small, distant license plates are hard: 44.8% of plates under 32 pixels leak (YuNet + YOLOv9-t: 51.3%).
Small faces leak more often than large ones: 17.6% under 32 pixels against 2.1% over 96.
Tiling finds small objects but costs time: on a CPU the tiled model is several times slower than YuNet + YOLOv9-t. Use the int8 model, or
--no-tileswhen speed matters more than small faces.Leak rates across skin tone, age, region and plate formats have not been measured yet.
No detector finds everything. For footage where a single leak matters, check the output, or run a second pass at a lower threshold.
Faces seen from behind carry little identity and are often not detected. Very small faces (under about 10 pixels) are missed more often.
Trained on photos from Open Images, which over-represent some regions. Plate formats that are rare there may be found less reliably.
Citation
@misc{ordnary2026meridianblur,
title = {Meridian Blur 1: face and license plate detection for privacy redaction},
author = {Ordnary},
year = {2026},
url = {https://huggingface.co/ordnaryai/meridian-blur}
}
Meridian™ is a trademark of Ordnary.