This work was a research driven project between RISE and Smart Eye, where we created an app to generate driver faces for training and validating Smart Eye’s Computer vision ML models.

Features

• Models tested: For synthetic data generation, we fine-tuned Generative Adverserial Networks and Latent Diffusion models with custom objective functions using Pytorch.
• Capacity of the model: The end result enabled engineers to specify which facial features, expressions, and ethnicity the image will contain.
• Use cases: The images were then used to train and validate in production ML models to prevent out of distribution and rare cases failures.

Impact

The app enabled the Smart Eye engineers to generate synthetic training and evaluation data for its ADAS system, thereby reducing cost and time of collecting data in natural settings, and reducing the risks and bias of its models.