After clicking “Watch Now” you will be prompted to login or join.
Generating Diverse and Photorealistic Synthetic Data for Real-World Perception Tasks
Nikita Jaipuria, Ford Motor Company | Rohan Bhasin, Ford Motor Company
GTC 2020
We'll cover these two related topics: First, leveraging generative adversarial networks for style transfer to diversify simulated images rendered in simple domains (that is, easier to render realistically, such as daytime) into photorealistic images in different weather and lighting conditions, using domain translation models (such as day-to-night, clear-to-rainy, clear-to-snowy) learned once from generic real-world datasets; and second, investigating the role of the discriminator's receptive field in unsupervised sim-to-real image translation. We'll show that reducing the discriminator's receptive field is directly proportional to improved structural coherence during translation in scenarios where the real and simulated images used for training have mismatched content — a situation often encountered in real-world deployment. Prior knowledge in computer vision and deep learning will help you get the most out of this session.