Case Study: Neurolabs improves synthetic data performance with Encord Active

Improving Synthetic Data Generation with Encord Active

Neurolabs, a computer vision company focused on in-store retail performance, needed a better way to train its models with synthetic data because performance on real-world data was lagging for certain classes. The team saw that synthetic references were not matching real-world conditions closely enough, leading to underperformance in edge cases and some classes scoring near zero.

Using Encord Active and its Quality metrics, Neurolabs analyzed model failures and data issues such as object scaling, pose, brightness, blur, and sample imbalance, then regenerated synthetic data to better reflect real-world examples. With Encord’s help, it improved overall last-mile performance by 2% and boosted one edge-case class by 67%, taking that class from 0% to 67% AP50 and from 0% to 100% P@1.

Neurolabs