Murd0ck/LicensePlateDetector_YOLOv8n

License Plate Detector (YOLOv8n)

Model Summary

YOLOv8n object detection model fine-tuned to detect Ukrainian vehicle license plates in images.

Model Details

Model Description

This model is a YOLOv8n-based object detector trained for license plate detection. It was fine-tuned from the Ultralytics YOLOv8n pretrained checkpoint.

• Developed by: WaveAutomaton
• Model type: Object detection (YOLOv8n)
• Framework: Ultralytics YOLO
• License: CC BY 4.0 (inherited from training dataset)
• Finetuned from model: yolov8n.pt

Model Sources

• Repository: local training run (not publicly hosted)
• Base model: https://github.com/ultralytics/ultralytics

---

Uses

Direct Use

License plate detection in images and video frames.

Downstream Use

• OCR pipelines
• traffic monitoring systems
• dataset preprocessing for recognition models

Out-of-Scope Use

• biometric identification
• surveillance beyond license plate localization without legal compliance

---

Bias, Risks, and Limitations

• Performance depends on lighting, angle, and region-specific plate formats
• May fail on heavily occluded or non-standard plates
• Not designed for identity inference beyond plate localization

Recommendations

Use with domain-specific validation. Do not assume cross-region generalization without testing.

---

How to Get Started

CLI

yolo detect predict model=best.pt source=images device=mps

Python

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("image.jpg")

for r in results:
    print(r.boxes)

Exported Models

• best.pt
• best.onnx

Training Details

Training Data

• AUTO.RIA Numberplate Options Dataset
• Source: https://nomeroff.net.ua/datasets/
• License: Creative Commons Attribution 4.0 (CC BY 4.0)

---

Training Procedure

• Model: yolov8n.pt
• Epochs: 154 / 350 (manual stop at convergence)
• Image size: 640
• Batch size: 16
• Device: Apple MPS
• Optimizer: auto
• Augmentation: default YOLOv8 (mosaic, HSV, flips)

---

Training Hyperparameters

• Mixed precision: AMP enabled
• LR scheduler: auto

---

Evaluation

Metrics (best validation epoch)

• mAP50: 0.9807
• mAP50-95: 0.8630
• Precision: 0.973
• Recall: 0.957

---

Testing Data

• Validation split defined in data.yaml

---

Technical Specifications

Model Architecture

YOLOv8n

---

Compute Infrastructure

• Apple Silicon (MPS backend)
• PyTorch 2.11
• Ultralytics 8.4.35

---

Environmental Impact

• Local training on Apple Silicon
• No cloud compute used

---