NGA SIBR Phase II Synthetic Data

In Phase II we will explore further what processing of synthetic image generation produces training data that optimizes the model’s performance on real world images. While existing work on CycleGAN optimizes for human visual perception of domain adaptation it may not necessarily be optimal for training a network to generalize to real scenarios. We will look at modifications of the loss functions used in GAN-based domain adaptation techniques to optimize for model performance instead of human perception. We will also look in more detail at what the differences are between real and synthetic images deep within the trained model to see why together they lead to better performance of the model than training on either set of data by itself.