AlessandroFerrante/StreetSignSenseY12m

StreetSignSenseYOLO12m

---

Model Summary

Street Sign Sense (YOLO12m) is an object detection model designed to identify and classify traffic signs in real-time. Based on the advanced YOLO12 Medium architecture, this model balances high accuracy with computational efficiency, making it suitable for Advanced Driver Assistance Systems (ADAS) research. It has been trained on the custom Street Sign Set, covering 63 distinct classes of traffic signs.

Usage

Live Demo

You can test this model instantly in your browser without any setup: 👉 Interactive Web Demo

Python

This model can be used with the Ultralytics framework or the official YOLO12 repository. It takes an image as input and outputs bounding boxes with class labels and confidence scores.

Code Snippet (Python)

from ultralytics import YOLO

# Load the model
model = YOLO('path/to/streetsignsense-yolo12m.pt')  # Replace with the downloaded model path

# Run inference on an image
results = model.predict(source='path/to/image.jpg')

# Show results
results[0].show()

Inputs: Images (RGB) of various resolutions (model trained at standard YOLO resolutions, e.g., 640x640). Outputs: List of Results objects containing bounding boxes (xyxy), class IDs, and confidence scores.

/StreetSignSenseY12m
├── .gitattributes   
├── README.md  
├── streetsignsense-yolo12m.pt   
└── metrics/ # metrics image folder

System

Standalone Model: Yes, this is a standalone object detection model. Input Requirements: Standard RGB images. No specific metadata required. Downstream Dependencies: The output (detected classes and locations) is intended to be used by decision-making logic in ADAS simulations or autonomous driving pipelines.

Implementation requirements

Hardware: Training was performed on Kaggle Notebooks using NVIDIA GPUs (e.g., Tesla P100 or T4). Software: PyTorch, Ultralytics YOLO framework. Compute:

• Training Time: 13h 23m 53s · GPU T4 x (depending on epochs).
• Inference: Capable of real-time performance (>30 FPS) on modern GPUs.

Model Characteristics

Model initialization

Fine-tuned: The model was initialized with pre-trained COCO weights (Transfer Learning) and then fine-tuned on the "Street Sign Sense" dataset to specialize in traffic sign detection.

Model stats

Architecture: YOLO12m (Medium). Characteristics: Utilizes attention-centric mechanisms to improve feature extraction compared to previous YOLO versions. Size: Medium-sized model, offering a trade-off between the speed of the 'Nano/Small' versions and the raw accuracy of the 'Large/X-Large' versions.

Other details

Precision: Trained using Mixed Precision (AMP). Pruning/Quantization: The uploaded weights are standard FP32/FP16. No post-training quantization has been applied yet.

Data Overview

Training data

The model was trained on the Street Sign Set (available on Kaggle).

• Source: A combination of public datasets and manually collected/annotated images.
• Size: Contains thousands of images with bounding box annotations.
• Classes: 63 specific traffic sign classes (speed limits, warnings, prohibitions, etc.).
• Preprocessing: Images were resized, and data augmentation (Mosaic, scaling, color adjustments) was applied during training to improve robustness.

Demographic groups

N/A: The dataset consists of street signs and environmental imagery. No human demographic data is involved or analyzed.

Evaluation data

The dataset was split into:

• Train: 70-80%
• Validation: 10-20%
• Test: 10% Differences: The test set contains unseen images from different environmental conditions to test generalization.

Evaluation Results

• Overview Risultati:
• Confusion Matrix:

Detailed Curves (Small)

Precision-Recall	F1 Score
Precision	Recall

Summary

The model achieves high Mean Average Precision (mAP) on the test set, demonstrating strong capabilities in detecting small objects (traffic signs at a distance) and operating in varied lighting conditions.

• Detailed Metrics: Please refer to the training graphs (F1-score, Precision-Recall curve) included in the attached notebooks.

Subgroup evaluation results

Performance is generally consistent across major classes (e.g., Speed Limits, Stop signs). However, classes with significantly fewer samples in the dataset may show slightly lower recall.

Fairness

Definition: Fairness in this context is defined as the model's ability to detect signs regardless of background clutter or slight occlusions. Results: The model shows robust performance in standard driving scenarios.

Usage limitations

• Lighting: Performance may degrade in extreme low-light conditions (night without streetlights) or heavy weather (dense fog/heavy rain) if not sufficiently represented in the training data.
• Occlusion: Signs that are more than 50% occluded may not be detected reliably.
• Geography: The model is trained primarily on European/International standard signs; it may not recognize signs specific to other regions that differ significantly in shape or color.

Ethics

Safety: This model is for research and educational purposes (ADAS development). It should not be used as the sole system for controlling a real vehicle on public roads without extensive safety validation and redundancy. Privacy: The dataset focuses on public street signs. Any incidental faces or license plates in the background are not the target of this model.

👨‍💻 Author

Alessandro Ferrante

Email: [email protected]