Rtdetr License Plate Detection Onnx

License-plate detector — RT-DETRv2 (R18), ONNX

On-device ONNX object detector that finds license plates in photos. A fine-tune of PekingU/rtdetr_v2_r18vd (Apache-2.0) on the Open Images "Vehicle registration plate" class, exported to a fixed 1×3×640×640 ONNX graph that runs fully offline with ONNX Runtime.

Built for KeepItLocal Privacy, a local-first redaction app, to suggest plate regions for blur/redaction — nothing leaves the device.

Intended use & scope

• Use it for: locating license-plate bounding boxes for privacy redaction.
• Suggestion, not a guarantee: treat each detection as a box a human confirms before redacting — like all detectors it can miss or over-fire.
• Low, well-ranked scores (important): this single-class fine-tune produces compressed confidence scores (the top box is typically ~0.10–0.15) but ranks them well — the highest-scoring box lands on the real plate ~92% of the time (mean IoU 0.74 on held-out data). Threshold low (~0.05) and take the top detection(s). This is an RT-DETR single-class trait, not a defect.

Files

File	Purpose
plate_rtdetr.onnx	the detector (fixed 1×3×640×640, ~79 MB)
config.json	RT-DETRv2 model config (1 class: license_plate)
preprocessor_config.json	the image-processor settings

I/O

• Input pixel_values: float32[1,3,640,640], RGB, scaled ×1/255, no mean/std normalization.
• Outputs logits: [1,300,1] (raw, apply sigmoid) and pred_boxes: [1,300,4] as cx,cy,w,h normalized to [0,1].

Usage (Python, ONNX Runtime)

pip install onnxruntime pillow numpy

import numpy as np, onnxruntime as ort
from PIL import Image

sess = ort.InferenceSession("plate_rtdetr.onnx", providers=["CPUExecutionProvider"])
img = Image.open("car.jpg").convert("RGB")
W, H = img.size
x = (np.asarray(img.resize((640, 640)), np.float32) / 255.0).transpose(2, 0, 1)[None]

logits, boxes = sess.run(None, {"pixel_values": x})     # [1,300,1], [1,300,4]
scores = 1.0 / (1.0 + np.exp(-logits[0, :, 0]))         # sigmoid
keep = scores > 0.05
for (cx, cy, w, h), s in zip(boxes[0][keep], scores[keep]):
    x0, y0 = (cx - w / 2) * W, (cy - h / 2) * H
    x1, y1 = (cx + w / 2) * W, (cy + h / 2) * H
    print(f"plate @ ({x0:.0f},{y0:.0f},{x1:.0f},{y1:.0f})  score={s:.2f}")

(RT-DETR is NMS-free; a light NMS on overlapping boxes is optional.)

Training

Fine-tuned from PekingU/rtdetr_v2_r18vd on ~3,000 Open Images plate images (the box-regression head kept its COCO pretraining; the classification head was re-initialized for the single license_plate class). Trained with the Hugging Face transformers Trainer; early-stopped where held-out loss bottomed (~epoch 8 of a 40-epoch run — later epochs overfit).

License & attribution

Apache-2.0, inherited from the base model.

• Base model: RT-DETRv2 (PekingU/rtdetr_v2_r18vd) — Apache-2.0
• Training data: Open Images V7, "Vehicle registration plate" class — images CC-BY-2.0 (Flickr), annotations CC-BY-4.0 (Google)