florence-research-chapter/vzc-traffic-light-dataset

VZC Traffic Light Dataset

VZC Traffic Light Dataset (VZC-TLD) is the accompanying dataset of the paper "Color is not enough: dataset and method for identifying relevant traffic lights in driving scenes" that will appear soon in IEEE Transactions on Intelligent Transportation Systems (T-ITS). The paper addresses a critical challenge for intelligent vehicles: determining which traffic lights are relevant for the ego-vehicle at intersections.

📊 Dataset Overview

The VZC Traffic Light Dataset is a COCO-format annotated dataset containing traffic light detections from a vehicle's perspective across diverse locations in the United States. The dataset distinguishes between relevant traffic lights (those that apply to the vehicle's current direction) and not relevant traffic lights (for other lanes/directions).

• Images: About 3,000 annotated images from 2,856 distinct intersections
• Traffic Lights: 25,624 distinct traffic light annotations
• Version: 1.0
• Year: 2025
• License: CC BY-NC 4.0
• Format: COCO JSON
• Image Size: 1280×720 pixels
• Coverage: Diverse US locations with varied weather and lighting conditions (day/night)

🚦 Traffic Light Categories

The dataset includes 8 distinct categories based on relevance and color state:

Relevant Traffic Lights

• relevant_unknown (ID: 4)
• relevant_green (ID: 5)
• relevant_yellow (ID: 6)
• relevant_red (ID: 7)

Not Relevant Traffic Lights

• not_relevant_unknown (ID: 0)
• not_relevant_green (ID: 1)
• not_relevant_yellow (ID: 2)
• not_relevant_red (ID: 3)

📁 Dataset Structure

VZC-traffic-light-dataset/
├── labels/                  # Dataset annotations
│   ├── train_v1.json     # Training set annotations
│   └── test_v1.json      # Test set annotations
├── img/                   # Image directory
│   └── *.png             # Dataset images (1280×720)
├── displayer.ipynb        # Interactive visualization tool
└── README.md              # This file

📓 Interactive Displayer Tool

The displayer.ipynb notebook provides a comprehensive visualization tool for exploring the VZC-TLD dataset. It allows you to:

• Browse through annotated images with color-coded bounding boxes
• Filter images by traffic light categories (relevant/not relevant, state)
• Visualize dataset statistics and distributions

🛠️ Installation & Requirements

The notebook automatically installs missing dependencies. Required packages:

• opencv-python - Image processing
• numpy - Numerical operations
• matplotlib - Visualization
• ipywidgets - Interactive widgets

🚀 Usage

Quick Start

1. Open displayer.ipynb in Jupyter or VS Code
2. Run all cells to load the dataset
3. Use the provided functions to explore images

Core Functions

Display Random Images

# Display random images from training set
display_random_images(train_data, "TRAIN", num_images=4)

Browse by Category

# Browse images with relevant red traffic lights (category_id=7)
browse_images_by_category(train_data, "TRAIN", category_id=7, max_images=5)

Interactive Browser Widget

# Browse all training images with slider control
create_image_browser_widget(train_data, "TRAIN")

📊 Annotation Format

The dataset follows the COCO format with bounding box annotations:

{
  "id": <annotation_id>,
  "image_id": <image_id>,
  "category_id": <0-7>,
  "bbox": [x, y, width, height],
  "area": <bounding_box_area>,
  "iscrowd": 0
}

📝 Citation

If you use this dataset, please cite the VZC Traffic Light Dataset paper:

@article{trinci2025color,
  title={Color is not enough: dataset and method for identifying relevant traffic lights in driving scenes},
  author={Trinci, Tomaso and Magistri, Simone and Bianconcini, Tommaso and Taccari, Leonardo and Sarti, Leonardo and Sambo, Francesco},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  year={2025},
  note={In Press},
}

📄 License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).

🤝 Contributing

For questions, issues, or contributions, please leave a message on this repository.