CCTV Traffic Scene Sematic Segmentation Dataset

Project Overview:

Objective

Our goal was to develop a comprehensive dataset derived from CCTV traffic scenes, specifically designed for semantic segmentation tasks in machine learning. This dataset is unique in its aim to precisely identify and classify each pixel in an image based on its associated object or area.

Scope

Collection of diverse traffic scenes from CCTV cameras and high-quality pixel-wise annotation to classify every segment of the image.

Sources

Urban intersections and traffic signals.
Highway traffic scenes.
Suburban streets.
Parking lots.
Pedestrian crossings.

Data Collection Metrics

Total Images: 100,000
Urban Intersections: 40,000
Highway Scenes: 30,000
Suburban Streets: 15,000
Parking Lots: 10,000
Pedestrian Crossings: 5,000

Annotation Process

Stages

Raw Image Pre-processing: Enhancing image quality, contrast, and removing noise.
Pixel-wise Labeling: Manual and semi-automated annotation tools used to mark each pixel.
Category Assignments: Pixels are categorized into classes like ‘Vehicle’, ‘Pedestrian’, ‘Road’, ‘Building’, ‘Tree’, etc.
Validation: Cross-verification of annotations by separate annotators.

Annotation Metrics

Total Annotations: 10 billion pixels (approximation)
Vehicle Pixels: 3 billion
Pedestrian Pixels: 1 billion
Road Pixels: 2.5 billion
Building Pixels: 2 billion
Tree Pixels: 1.5 billion

Quality Assurance

Automated Segmentation Models: Used pre-trained models to cross-verify manual annotations.
Expert Review: A team of experts reviewed challenging segments for accuracy.
Inter-annotator Agreement: Several annotators reviewed the same image, ensuring pixel-wise consistency.

QA Metrics

Annotations Reviewed by Experts: 20 million pixels (0.2% of total annotations)
Inconsistencies Identified and Rectified: 5 million pixels (0.05% of total annotations)

Conclusion

The CCTV Traffic Scene Semantic Segmentation Dataset has been meticulously crafted to ensure high precision in every pixel. It offers a comprehensive view of various traffic scenes, ensuring versatility and depth for researchers and developers in the field of computer vision.