Model card · Version 1.0 · Released November 2026
Meridian™ Blur 1 model card
| Model page | Meridian™ Blur |
|---|---|
| Resources and technical documentation | Hugging Face: ordnaryai/meridian-blur |
| Terms of use | Apache License 2.0, and Ordnary's AI Usage Policy |
| Authors | Ordnary |
Model information
Summary description and brief definition of inputs and outputs.
Description
Meridian Blur 1 is an object detection model that finds human faces and vehicle license plates in images, so they can be made unrecognizable before footage is stored, shared or published. It is built and trained by Ordnary on public data, released with open weights under the Apache 2.0 license, and small enough to run on an ordinary CPU.
Inputs and outputs
- Input: an RGB image of any size. It is resized to the model's input resolution; images much larger than that are also scanned in overlapping tiles, so small faces and plates stay visible to the model. Video is processed frame by frame.
- Output: a list of detections, each with a class (
0face,1license_plate), a bounding box in pixels of the original image, and a confidence score between 0 and 1.
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}
} Model data
Data used for model training and how the data was processed.
Training dataset
Meridian Blur 1 is trained on images from Open Images V7 that contain the classes "Human face" or "Vehicle registration plate". The annotations are licensed CC BY 4.0 by Google; the images are licensed CC BY 2.0 by their authors. The exact list of images used is published with the model.
No images of Ordnary customers, and no images collected without a license that allows this use, are part of the training data.
Data preprocessing
- Only the two classes are kept; "Human face" becomes face, "Vehicle registration plate" becomes license_plate.
- Images are rotated upright using the published rotation data, and boxes converted to pixel coordinates.
- The Open Images validation and test images are kept apart and never used for training; they form the evaluation set.
Implementation information
Details about the model internals.
Hardware
The model was trained on NVIDIA data-center GPUs. It runs on any x86 or ARM CPU and on GPUs through ONNX Runtime; no GPU is needed for photos or recorded video.
Software
The architecture is RF-DETR, a real-time detection transformer released under Apache 2.0, fine-tuned with PyTorch. The released weights come as ONNX for inference and safetensors for further training. A single Python script, inference.py, blurs an image or a video and needs only ONNX Runtime, OpenCV and NumPy.
Evaluation
Model evaluation metrics and results.
Benchmark results
Meridian Blur 1 is evaluated on 14,431 held-out Open Images V7 images (3,639 from validation, 10,792 from test) with 21,251 labeled faces and 3,808 labeled license plates. None of these images were used for training or for choosing a checkpoint, and none carry pseudo-labels. Every model below ran on the same images and was scored by the same code. Meridian runs at a score threshold of 0.3, with the full image plus 960 pixel tiles; the other models run at their own default thresholds and input sizes.
Leak rate: share of labeled objects less than 60% covered by the redaction, with the same padding for every model (35% for faces, 12% for plates). By size, on the long side of the labeled box: small under 32 pixels, medium 32 to 96, large over 96. Lower is better. mAP50: average precision at an overlap of 50%, measured at a low threshold so every model has a score range.
Faces
| Model | Threshold | Leak, all | Small | Medium | Large | mAP50 |
|---|---|---|---|---|---|---|
| Meridian Blur 1 | 0.3 | 3.6% | 17.6% | 4.1% | 2.1% | 0.903 |
| Meridian Blur 1, int8 | 0.3 | 3.7% | 18.6% | 4.3% | 2.2% | 0.903 |
| Meridian Blur 1, full image only | 0.3 | 4.8% | 23.3% | 6.0% | 2.7% | 0.907 |
| CenterFace (deface 1.5.0) | 0.2 | 10.8% | 16.3% | 9.6% | 10.9% | 0.784 |
| RetinaFace ResNet-50 | 0.6 | 11.8% | 35.5% | 13.4% | 9.0% | 0.833 |
| YuNet (OpenCV Zoo, 2023) | 0.5 | 15.5% | 26.0% | 15.0% | 14.9% | 0.698 |
| MediaPipe face detection, full range | 0.5 | 47.6% | 97.6% | 73.7% | 31.1% | 0.443 |
License plates
| Model | Threshold | Leak, all | Small | Medium | Large | mAP50 |
|---|---|---|---|---|---|---|
| Meridian Blur 1 | 0.3 | 10.6% | 44.8% | 5.3% | 1.5% | 0.857 |
| Meridian Blur 1, int8 | 0.3 | 10.9% | 45.8% | 5.6% | 1.5% | 0.858 |
| Meridian Blur 1, full image only | 0.3 | 12.4% | 50.9% | 7.1% | 1.9% | 0.871 |
| YOLOv9-s 608, with 960 px tiles | 0.25 | 9.3% | 43.7% | 3.7% | 0.4% | 0.843 |
| YOLOv9-s 608 | 0.25 | 11.7% | 54.9% | 4.6% | 0.4% | 0.866 |
| YOLOv9-t 640, with 960 px tiles | 0.25 | 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, and also without tiles. CenterFace leaks fewer of the smallest faces, partly because it marks many more boxes (its share of detections that match a labeled face is 69%, against 80% for Meridian). On license plates the models are level: the YOLOv9-s plate model with tiles leaks less (9.3% against 10.6%) but more of what it blurs is not a labeled plate (75% of its detections match one, against 90%). MediaPipe is built for faces close to the camera and sees the image at 192 pixels, so it finds almost no small faces. Meta's EgoBlur was not tested: its weights are only available after accepting a license form. The YOLO plate models are from open-image-models; CenterFace is the model used by the open-source tool deface.
Latency: median time per image on a 4-core CPU, over 15 runs. Tiling finds small faces and distant plates but costs time; it can be turned off.
| Model | 1280 px | 2400 px |
|---|---|---|
| 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, tiled | 137 ms | 330 ms |
| CenterFace (deface) | 388 ms | 3,416 ms |
| RetinaFace ResNet-50 | 902 ms | 3,348 ms |
| YOLOv9-s, tiled (plates only) | 393 ms | 921 ms |
| MediaPipe (faces only) | 10 ms | 18 ms |
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. GPU latency has not been measured yet.
Ethics and safety
Ethics and safety evaluation approach and results.
Evaluation approach
- Bias: the leak rate is measured separately for small, medium and large faces, and on images from different regions, to find groups of people the model protects less well.
- Privacy: the training data is public, licensed for this use and listed in full. The model runs where the footage is, so images never have to be sent to Ordnary.
- Safety and security: the model is tuned to over-blur rather than miss, and MeriJudge checks redacted frames for anything still recognizable.
- Explainability: every decision is a box with a score that can be drawn on the image, and the thresholds and padding are published.
Evaluation results
The leak rate by object size is in the benchmark results above. The model protects small faces least well: 17.6% of faces under 32 pixels leak, against 2.1% of large faces. Leak rates across skin tone, age, region and license plate formats have not been measured yet. It is the next evaluation, and the results will be published here.
Usage and limitations
These models have certain limitations that users should be aware of.
Intended usage
- Blurring faces and license plates in photos, dashcam footage and camera streams before they are stored, shared or published.
- Privacy by design in products that handle images of the public, under the GDPR and similar laws.
- Preparing image datasets for research or training without identifiable people.
Limitations
- Faces smaller than about 10 pixels are missed more often, and faces seen from behind are often not detected.
- Plates in direct headlight glare, heavy rain or strong motion blur can be missed.
- Plate formats that are rare in Open Images may be found less reliably.
- Infrared, fisheye and very dark footage were not part of training.
- No detector finds everything. Where a single leak matters, the output should be checked.
Ethical considerations and risks
A detector that finds faces could be misused to find people. Meridian Blur 1 only says where a face or a plate is, so that it can be hidden; it does not recognize who someone is or read plate numbers. Using it to identify, track or profile people, or as part of a surveillance system, is outside its intended use and against Ordnary's AI Usage Policy.
Benefits
Meridian Blur 1 gives everyone who handles images of the public an open, inspectable and freely usable way to protect the people in them, without sending footage to a third party.