Registration Plate Detector

This repository provides a complete pipeline for detecting vehicle registration plates using Ultralytics’ YOLOv11 framework. It automates downloading, unzipping, and converting the raw dataset into YOLO‐formatted train/validation splits with normalized annotations. The notebook demonstrates visualizing ground-truth bounding boxes, training and validating a YOLOv11s model, and evaluating its performance using COCO metrics. Finally, it includes helper functions for running inference on both images and videos, producing annotated outputs to showcase real-time plate detection.

---

Example Results

Below are a few sample predictions from the trained model, showcasing its performance on validation images. Each image displays the predicted label with confidence and the true label for comparison.

Sample Predictions:

Inference over static images

Example 1

Example 2

Example 3

Inference over a video

Training Performance Training and Validation Dataset:

Training Loss/Acc on Validation and Training Dataset

---

What it does

• Dataset Preparation: Automatically downloads and extracts the raw vehicle registration plate dataset, then converts the annotations into YOLO-compatible format with normalized bounding boxes.

• Data Organization: Creates a structured directory hierarchy (train/, valid/, labels/, images/) and sets up symlinks for easy training and validation splits.

• Visualization: Samples images at random to display ground-truth bounding boxes and, later, side-by-side comparisons of predictions vs. annotations.

• Training & Validation: Generates a YAML configuration for Ultralytics’ YOLOv11s model, kicks off training for a specified number of epochs, and evaluates performance on the held-out validation set.

• Evaluation: Converts predictions and ground-truths into COCO format and runs the standard COCO evaluation script to compute AP metrics.

• Inference: Provides helper functions to run detection on individual images or entire videos, drawing and saving annotated outputs for real-time plate detection demos.

---

Run Instructions

Clone the repository and build all projects in one step:

git clone https://github.com/sancho11/registration_plate_detector.git
cd registration_plate_detector
python -m venv .venv
source .venv/bin/activate  # On Windows, use: .venv\Scripts\activate
pip install -r requirements.txt

To run the project:

#For running the notebook pipeline using jupyter notebook
jupyter notebook
#For training a model using python
python train.py
#For running evaluation metrics on a trained model
python evaluate.py
#For running classification on a single image.
python infierence.py

Training

python train.py --epochs 100 --batch 16

Evaluating

Generate confusion matrix and do inference over whole dataset

python evaluate.py --model model/path/checkpoint --output-dir results/directory/path

Inferring

Do an inference on a single image using a trained model

python infer.py --model path/to/model /path/to/image/or/video output/file/path/name

---

Pipeline Overview

Pipeline Diagram

---

Key Techniques & Notes

• Annotation Conversion: Uses a custom two_corner_boxes_to_normalized_bbox function to convert raw corner-based annotations into YOLO’s [x_center, y_center, width, height] format, ensuring six-decimal precision for stable training.
• Symlinked Data Splits: Leverages filesystem symlinks to avoid duplicating large image folders, while organizing train/ and valid/ directories under a single yolo_dataset/ root.
• Ultralytics YOLO Integration: Generates a minimal YAML config (nc, names, paths), toggles TensorBoard logging via !yolo settings tensorboard=True, and executes yolo task=detect mode=train … for streamlined training.
• Visualization Utilities: Implements plot_box, show_random_samples, and show_random_predictions to render ground-truth and predicted boxes side by side using OpenCV and Matplotlib, aiding qualitative analysis.
• COCO-Style Evaluation: Converts annotations and detections into COCO JSON format with generate_dataset_and_preditions_cocoeval_format, then runs pycocotools.COC Oeval to report standard AP metrics.
• Video Inference Pipeline: Defines video_read_write to iterate over video frames, apply YOLO detection, draw boxes, and write annotated output—demonstrating real-time plate detection on arbitrary video files.
• Modular, Well-Documented Code: All steps are wrapped in self-contained functions with clear docstrings, making it easy to adapt or extend for different datasets or model architectures.