DETR (License Plate Detector): spotting plates in traffic photos, and where to run it

Two things to settle before anything else. This model is under legal review and is not available to users in the European Union or the United Kingdom; everywhere else it runs in your browser on CNAPS Studio, included in the basic plan. And its licence is not one of the standard open ones. The author, Nicholas Muchinguri, granted commercial use directly with attribution required, and that grant is what we operate under. Self-hosting means relying on the same permission rather than a public licence file. At 41.6 million parameters the technical side is easy: PyTorch, the Hugging Face transformers library, and a GPU only if you want speed.

The job is narrow and useful. Give it a photo of traffic, a dashcam frame, or a parking-lot camera shot, and it draws a box around every licence plate it can see, with a confidence score for each. It is a fine-tune of the general DETR detector with the classifier reduced to a single class.

The cost is that there is nothing published to check it against. No benchmark numbers, no evaluation, not even a model card. Two DETR fine-tunes sit in the CNAPS Studio catalog and this is the one that reads vehicles rather than people, on the understanding that its quality is something you establish yourself.

What it actually is

The architecture comes over from DETR untouched: a convolutional backbone reads the image, a six-layer transformer encoder builds a view of the whole frame, a decoder runs 100 object queries against it, and the heads turn each query into a class and a box. The classifier has one class, and the 100-query ceiling caps an image at 100 plates, which no realistic photo will reach.

It was trained from the general DETR checkpoint on licence-plate data covering a range of plate formats and countries. The exact dataset is not named anywhere public, so which countries are well covered is unknown, and that is a real question for anyone outside the places most such datasets are collected.

Output is a list. For each detection: the label licence-plates, a box given as left, top, right and bottom coordinates, and a confidence score from 0 to 1. Plate detection is only ever step one, because a box is not a plate number.

What is published, and what is not

This is the shortest version of this section in the series, because the Hugging Face page carries no model card at all. No accuracy score, no evaluation set, no intended-use notes, no limitations, no licence file. What is published is the parameter count of 41.6 million, the base checkpoint it came from, and 91 downloads in the last month.

Our own documentation is honest about the same gap and offers qualitative notes instead: clear frontal plates are the strong case, accuracy falls as the angle steepens, severe angles and occlusion cause misses, several plates in one photo are handled well, daylight is best and night works with adequate lighting, and small or distant plates may be missed. Those are descriptions rather than measurements, and we wrote them.

So the burden of proof is entirely on you. Before this goes near billing, enforcement, or an access barrier, run a few hundred frames from the actual camera at the actual angle in the actual light and count the misses. That is true of any detector, and doubly true of one with no published evaluation.

How it compares to the other detectors here

There is no honest numeric comparison to make. This model publishes no score, so putting it in a table next to DETR's 42.0 or RF-DETR Medium's 54.7 would create a ranking out of one real number and one blank.

What separates them is what they look for. Reach for this when the target is specifically a plate, since a single-class detector will not confuse one with a road sign. Reach for the general DETR or RF-DETR Medium when you want the vehicle rather than the plate, since car, truck and bus are on their category lists and both publish accuracy figures. Reach for SAM 3.1 (Scene) when you want the plate as a mask rather than a box, and note it is not region-restricted.

What to chain it with

It takes one image and returns labelled boxes with confidence scores, which is exactly the input the rest of the catalog expects from a detector.

Two chains do the real work. For reading plates, Image Crop by Coord cuts each box out and PaddleOCR reads the characters from the crop, which is the whole automatic plate-reading pipeline in two steps. For the opposite job, Image Masker by Class turns the boxes into a mask and Selective Blur obscures the plates, which is what you want before publishing street photography or dashcam footage. For video, run Video Split first and Video Reassemble afterwards.

Open DETR (License Plate Detector) in CNAPS Studio, feed it ten frames from your own camera at your own mounting angle, and count the plates it missed before you build the rest.

Sources

• huggingface.co/nickmuchi/detr-resnet50-license-plate-detection for the 41.6 million parameter count, the base checkpoint, the monthly download count, and the absence of a model card, licence file, evaluation numbers and limitations section.
• huggingface.co/facebook/detr-resnet-50 for the inherited architecture, the 100 object queries, and the 42.0 average precision quoted above.
• docs.cnaps.ai/ai-model-hub for the CNAPS Studio input and output, the author-granted licence terms, the qualitative accuracy notes, and the regional restriction and legal review status.