Scania Digit Detection (Hotspot Detection)

Lightweight notebook and scripts for detecting and grouping characters (digits/letters) on Scania images using YOLOv8.

Contents

• hotspot_detection_scania.ipynb — main notebook: data conversion, training, prediction, grouping, visualization, and XML generation.
• new_dataset/ — sample dataset structure (images + labels + data.yaml).

Prerequisites

• Python 3.8+
• macOS (tested) or Linux

Installation

Run:

pip install -r requirements.txt
# or
pip install ultralytics opencv-python matplotlib tqdm pyyaml

(If you prefer to use the notebook cell, it already contains !pip install ultralytics opencv-python matplotlib.)

Quick Start

1. Open hotspot_detection_scania.ipynb in Jupyter/VS Code.
2. Edit configuration placeholders in the notebook cells (paths such as image folder, xml folder, model paths, dataset folder).
3. Run the cells in order:
   • Install dependencies
   • Convert Pascal VOC XMLs to YOLO labels (if needed)
   • Organize dataset into train/val and create data.yaml
   • Train YOLOv8
   • Visualize training metrics
   • Run predictions and group characters
   • Generate XML annotations from predictions

Dataset layout (expected)

new_dataset/

• images/
  • train/
  • val/
• labels/
  • train/
  • val/
• data.yaml

YOLO label files are .txt with class and normalized bbox values.

Training

The notebook uses Ultralitycs YOLO API (YOLOv8). Example training cell:

from ultralytics import YOLO
model = YOLO('yolov8n.pt')
model.train(data='new_dataset/data.yaml', epochs=100, imgsz=640, batch=20, device='mps')

Adjust device, epochs, batch, and imgsz as required.

Inference & Grouping

Prediction cells show how detected character boxes are collected, sorted, grouped (left-to-right/top-to-bottom) and merged into multi-character boxes. There are two grouping variants in the notebook: a simple y/x sort method and an improved xyxy gap-based method.

XML Generation

Notebook includes functions to convert grouped detections into Pascal VOC XML files for downstream use.

Notes

• Replace placeholder paths in the notebook before running.
• Ensure model paths point to existing weights (best.pt).
• The grouping thresholds may need tuning per image set.