Autonomous Driving Perception

Multi-Class Image & Video Annotation

Project Overview

Powering the next generation of Smart Mobility through high-precision computer vision datasets and ground-truth labeling.

In the realm of Autonomous Vehicles (AV), the difference between safety and failure lies in the quality of the training data. Zyntriq was tasked with developing a high-fidelity dataset for an autonomous driving perception model. This project involved complex multi-class object detection and spatial tracking across both high-resolution static images and dynamic urban traffic video streams.

The Challenge

Raw visual data from vehicle-mounted cameras is often cluttered, overlapping, and subject to varying lighting conditions. The client required:

  • Extreme Precision: Pixel-perfect bounding boxes for objects at varying distances.
  • Temporal Consistency: Maintaining object IDs across video frames to assist in motion prediction.
  • Multi-Class Complexity: Simultaneous labeling of diverse categories including vehicular traffic, vulnerable road users, and infrastructure signals.

The Zyntriq Solution

Our team deployed a rigorous multi-stage annotation pipeline designed to eliminate human error and maximize model performance.

1. Comprehensive Multi-Class Labeling

We categorized and labeled four critical classes essential for urban navigation:

  • Vehicles: Covering cars, trucks, motorcycles, and buses with strict adherence to occlusion rules.
  • Pedestrians: High-detail tracking of human movement to support collision avoidance systems.
  • Traffic Lights: Identifying signal states (Red, Yellow, Green) and positioning to aid in decision-making logic.
  • License Plates: High-zoom annotation for vehicle identification and security training.

2. Video Tracking & Interpolation

Utilizing CVAT’s advanced tracking features, we ensured that moving objects remained consistent across frames, providing the temporal data necessary for the AI to understand velocity and trajectory.

3. Quality Assurance & Dataset Curation

Every task underwent a "Two-Pass" verification system. Our senior annotators audited the datasets to ensure 99.9% accuracy in bounding box placement and attribute tagging before exporting to the final training environment.

Key Technical Capabilities Demonstrated

  • High-Resolution Zoom Annotation: Ensuring clarity on small or distant objects like license plates and traffic signals.
  • Occlusion Handling: Expertly labeling partially hidden objects to help models recognize "hidden" hazards.
  • Dataset Interoperability: Converting raw annotations into industry-standard formats (COCO, XML) for immediate integration into the client’s deep learning frameworks.

The Result

By delivering a robust, clean, and accurately labeled dataset, Zyntriq enabled the client to:

  • Reduce Model Training Time: High-quality ground truth data led to faster convergence during the training phase.
  • Improve Perception Accuracy: The model saw a significant percentage increase in the detection of pedestrians and traffic signals in low-light urban environments.
  • Scalable Workflow: Developed a repeatable framework for processing thousands of hours of road footage with consistent quality.

Industry

Automotive / AI & Machine Learning

Service

Image & Video Annotation, Dataset Curation

Tools Used

CVAT (Computer Vision Annotation Tool)

Output Formats

COCO 1.0, CVAT XML, YOLO


Resources

View Project Files

Ready to Scale Globally?

Does your AI model need high-precision training data? Zyntriq provides the human-in-the-loop expertise needed to scale your computer vision projects.