Published: 12 April 2023
Summary
Corner edge cases coverage and operational cost in driving millions of miles to collect real data, plus labeling it for training AI algorithms, are top challenges for autonomous vehicle proliferation. Product leaders must leverage street-scene synthetic image data for validating and testing AVs.
Included in Full Research
Overview
Key Findings
Techniques to generate synthetic image data are continuously evolving and relying solely on any specific technique like game engine, generative adversarial networks (GANs) or computer simulations in isolation for generating synthetic image data will have domain gaps.
While synthetic data provides a shortcut to real data, there are drawbacks in solely relying on it, such as overfitting to situations that might never happen and the confidence of coverage for real-world scenarios.
With synthetic data, random noise can be interjected to simulate dirty cameras, fog and other visual obstructions, providing an avenue to cover edge cases that constitute <1% of real-world
Clients can log in to view the entire
document.