Licence Plate Detection

Project overview

Property and real estate images were annotated for a computer vision dataset designed to detect vehicle licence plates for privacy protection. Accurate four-point polygon annotations were created to generate high-quality training data for an AI model capable of automatically detecting and blurring licence plates before property images are published, replacing a third-party detector prone to false positives.

Work performed

• Reviewed and organised the property image dataset before annotation.
• Accurate four-point polygon annotations were created around every visible licence plate.
• A consistent corner convention, starting bottom-left and running upward, made every polygon machine-comparable.
• Images containing no plates were confirmed and marked as verified-empty negatives.
• Annotation consistency was maintained across all images and rounds.
• Quality assurance checks included a geometry audit that flagged and corrected anomalous polygons before delivery.
• Completed annotations were delivered in COCO 1.0 format according to client requirements.

What made it hard

• Annotating extremely small plates, with a median plate width of 91 px on 4,000 px images and the smallest at 12 px.
• Rotated quadrilaterals rather than axis-aligned boxes were needed so angled plates are enclosed tightly without including bumper and paintwork.
• Genuine plates had to be distinguished from phone numbers on van signage, permit stickers, house numbers and reflective badges.
• Plates that were partially occluded, clipped by the image edge, or seen at extreme angles required careful handling.
• A consistent corner ordering across thousands of polygons was needed to keep the data usable for key point training.
• Disputed annotations where the model and the original labels disagreed called for repeat review passes.

Result

A high-quality annotated dataset was delivered for privacy-aware computer vision applications. The dataset was used to train licence plate detection models that reduced missed plates from 39 to 4 on a held-out test set, improving detection against the client's deployed system.

Final deliverable

Annotation of 2,328 property images was completed, comprising 3,267 licence plate polygons and 584 verified-empty negatives, using a single plate class annotated as four-point rotated quadrilaterals with a consistent corner convention. Delivery spanned three annotation rounds with repeat quality-assurance passes, and the dataset was prepared for training an AI-powered automatic licence plate blurring system.