The ICC uses machine learning (ML) and computer vision extensively to develop products that improve over time and with customer usage. One of the key components of these ML-powered products is structured and labelled data. Acquiring this data in the real world can be time-consuming and expensive. At the ICC, synthetic data is emerging as a more affordable, scalable alternative to generating this ML training data.
“Most perception neural networks rely on labeled data, which is costly and prone to error,” said the ICC’s Elnaz Vahedforough, technical project manager. “By using synthetic data, once the labeling task is set up, the labeling is essentially free, and other costs are minimized.”
The ICC generates images and ground-truth data in Unity to train neural networks for implementation of autonomous driving components, such as sensors, perception, prediction and driving.
Besides the cost considerations, synthetic data generated with Unity can be used to construct scenarios that rarely occur (e.g., accidents, unusual objects on the road, etc.) or harsh weather conditions such as fog or heavy rain. Vahedforough noted, “This makes it possible to recreate edge-case scenarios safely.”
Source: Unity Technologies Blog