July 08 2021
How TRI Trains Better Computer Vision Models with PD Synthetic Data
Written by Kevin McNamara, Founder & CEO
We recently had the opportunity to sit down with Toyota Research Institute’s Head of Machine Learning Research, Adrien Gaidon, to:
- Discuss TRI’s most recently published research that utilizes Parallel Domain’s synthetic data to beat state-of-the-art results across a variety of tasks, such as object tracking, depth estimation, and semi-supervised semantic segmentation.
- Recap our collaborative efforts to minimize the domain gap between synthetic and real data, resulting in models that learn on our synthetic data and can adapt to the real world.
- Get his answer to the question, “Does synthetic data work?” Spoiler alert, the answer is yes!
“When it comes to object tracking, we improved the state of the art. We improved the robustness. It’s not just accuracy, it’s also safety benefits that directly resulted from combining synthetic data with real world data. Always mixing the two is the best of both worlds.” – Adrien Gaidon, Head of Machine Learning Research, Toyota Research Institute
Be sure to also check out TRI’s research papers that use Parallel Domain’s synthetic data to beat state-of-the-art results across several different benchmarks:
- Learning To Track with Object Permanence [arXiv link]
- Geometric Unsupervised Domain Adaptation for Semantic Segmentation [arXiv link]
Check out our Data Visualizer to easily view samples of our synthetic data for autonomy applications!