Scaling Leading Automotive AI Model Development with End-to-End Data Annotation
.png){kind=logo}
The Challenge
As leaders in the automotive AI industry, this customer's goal is to deliver the highest quality AI models, which consistently improve at a faster rate than their competitors. They’re developing models to recognize vehicle type, make, model, as well as OCR number plate reading (across varying national and international formats) for use cases including automated parking detection.
One key challenge for developing automotive recognition models at this scale was creating a global dataset of different national vehicle makes, and license plate formats for >99.9% model accuracy in the production environment.
A dedicated team for continuous training data production was needed to ensure the high data throughput requirements didn’t affect data accuracy or hinder speedy development timeline.
The Solution
Outsourcing these labeling challenges to a dedicated annotation team from Digital Divide Data through V7 Darwin has massively streamlined data production.
High quality annotations and data reviews before routing back to the customer eliminated the hours previously required to review and filter datasets, resulting in significant time savings, and improved data coverage for better performing models.
Training global datasets at speed
To train their automotive AI models for global applications, this team needed to annotate a vast quantity of internationally routed data, up to 200,000 images a week, extremely efficiently. This was made possible by V7’s scalable dataset management features, seamless external storage integrations, and AutoML labeling tools.
This project demanded an annotation team both familiar with the V7 platform and able to specialize in recognising different license plate formats (between US states, and on an international level) at speed, which DDD was able to deliver.
V7 and Digital Divide Data achieve over 99.9% annotation accuracy
When searching for the best tools for a data annotation project of this scale, this team tried several other providers before selecting our joint solution with Digital Divide Data.
Our joint offering combined V7’s best-in-class annotation and QA tools, with Digital Divide Data’s team of labeling experts to deliver the highest quality training data.
Their key priorities included annotation accuracy, data throughput speed, and constant communication on project progression. They were very impressed with the communication and level of attention they received from the DDD team, and the labeling accuracy they were able to achieve on the V7 Platform.
As a result of this partnership, the DDD team was able to quickly ramp up the project from 35,000 images a week to around 200,000 images - helping the customer accelerate their model development, and reach goals of out-competing competitors. As well as achieve over 99.9% accuracy in their training data using only a small team within DDD’s scalable workforce.
Learn more about our partnership with Digital Divide Data here.
