How Hugo Supported UC Berkeley’s Autonomous Vehicle Project

The Client

UC Berkeley’s DeepDrive (BDD) is a leading research group focused on autonomous driving and developing machine learning systems for self-driving vehicles. Their mission is to develop AI that matches and surpasses human visual processing and decision-making abilities on the road.

The Challenge: Bridging the Visual Intelligence Gap

Despite significant advancements, autonomous driving AI still lags behind human capabilities in processing complex visual relationships in real-world scenarios. This gap poses risks in ensuring safe navigation and timely decision-making, which are critical for successfully deploying autonomous vehicles. To address this challenge, BDD needed to:

1. Annotate a vast dataset comprising millions of images, videos, and sensor outputs to train and validate automotive ML models.
2. Rigorously benchmark their in-house data labeling tool against industry standards for accuracy, quality control, and efficiency.

“The current state of self-driving AI can’t match human drivers’ contextual awareness and quick decision-making,” explained BDD’s Head of Machine Learning. “To overcome this, we need extensive labeling across countless real-world scenarios.”

Hugo’s Solution

Hugo developed a comprehensive approach across four major areas to support BDD’s mission:

Assembled Specialized Annotation Experts

Hugo carefully curated a team of domain experts, all STEM graduates with diverse backgrounds in mechanical engineering and automotive design. This ensured the annotations were not only accurate but also contextually relevant, directly addressing the challenge of complex visual scenarios.

Comprehensive Annotation Capabilities

Hugo’s team of annotators delivered precise annotations across various modalities, including;

• 2D-bounding boxes
• Pixel-level semantic segmentation
• 3D point cloud processing
• Multi-sensor fusion and data labeling

These diverse annotation modalities captured both simple and complex visual relationships, ensuring a thorough understanding of the data.

Rigorous Quality Control

To enhance consistency and accuracy, Hugo improved BDD’s annotation guidelines to ensure clarity and comprehensiveness. This enabled the team to achieve quality scores exceeding 98% and effectively address quality control challenges in real-world scenarios, such as low-light video feeds and densely packed frames.

Workflow Optimization

Hugo engineered a batching workflow that facilitated systematic feedback on failure modes and opportunities for improvement, enabling continuous improvement of BDD’s internal data labeling pipeline.

The Results: Accelerating Progress with Speed and Accuracy

Hugo’s expertise accelerated BDD’s development of autonomous vehicles:

• Annotated 4,000 images and videos in under 1 month—89% faster than BDD’s previous vendor average.
• Achieved annotation quality 3 percentage points higher than BDD’s target of 95%.
• Comparative analyses identified 11 critical fail points in BDD’s data labeling tool.
• Annotation consistency scores regularly exceeded 98%, even on challenging nighttime and high-density inputs.

“On the most difficult edge cases that tripped up other vendors, Hugo’s team delivered with flying colors,” remarked the Head of Machine Learning.

A Valuable Partnership

By delivering accurately annotated data at scale, Hugo became a crucial partner in supporting BDD’s pursuit of autonomous driving AI. “Hugo’s ability to scale teams of specialized annotators has been crucial,” the Head of Machine Learning concluded. “Their support has allowed our scientists to focus on core research, accelerating our progress towards autonomous vehicle integration.”