AlessandroFerrante/StreetSignSet

Street Sign Set 🚦

This dataset was developed for the Street Sign Sense project, which achieved a final grade of 30/30 in the Machine Learning course. It is specifically designed and optimized for training state-of-the-art object detection models, such as YOLOv12, for the recognition of Italian and European road signs.

Dataset Description

The Street Sign Set is a comprehensive collection of traffic sign images, annotated in the YOLO format. It focuses on real-world driving conditions to ensure robustness in Advanced Driver Assistance Systems (ADAS).

Street Sign Set is a dataset with over 7300 images designed for road sign detection in realistic contexts.

🏷️ Class Structure and Labels

The dataset comprises 63 total classes, organized into 5 macro-categories that define the label prefix: prio (priority), forb (prohibition), info (information), warn (warning), mand (mandatory).

Labeling examples: prio_give_way, forb_speed_over_50, info_parking, warn_right_curve, mand_pass_left_right.

It includes 23 main classes identified as primary targets, including:

• Speed limits: 14 classes (e.g., 5–130 km/h).
• Prohibition signs: 4 classes (e.g., no stopping and parking, no overtaking).
• Priority signs: 2 classes (e.g., give way, stop).
• Curves and crossings: 3 classes (e.g., dangerous curves, pedestrian crossing).

🛠️ Hybrid Origin and Construction

• Base: ~4000 images from a dataset available on Kaggle.
• Expansion: ~3000 images manually integrated from external sources and street mapping services to cover underrepresented classes, subsequently manually labeled.

The dataset is not perfectly balanced due to the frequency of road signs in reality, so some signs appear more often than others.

⚙️ Technical Specifications

• Format: Standard YOLO annotations (.txt).
• Filename: Rigorous logical scheme class_name-n.jpg (e.g., prio_give_way-12.jpg).
• Selective Data Augmentation: Applied only to rare classes to mitigate imbalance. It includes variations in Hue/Saturation/Brightness, Grayscale (23%), Blur, and Noise to simulate adverse conditions.

Dataset Structure

The data is organized following the standard YOLO convention, making it ready for immediate training:

.
├── train/
│ ├── images/ # Training set
│ └── labels/ # YOLO annotations
├── val/
│ ├── images/ # Validation set
│ └── labels/ # YOLO annotations
├── test/
│ ├── images/ # Test set for final evaluation
│ └── labels/ # YOLO annotations
├── data.yaml # Dataset configuration file (classes names)
└── dataset_analysis.csv # Detailed analysis of the dataset class distribution