Synthetic data is artificially generated data created by scientific simulators, rendered environments, generative AI systems, or other synthetic sources. It can help bootstrap an ML project when suitable real-world data is scarce, but usefulness depends on how closely it reflects deployment conditions. section 6.1 section 6.5
Use synthetic data with practical checks:
- Treat straightforward rendering pipelines cautiously: imperfect lighting, shadows, and other realism gaps can make synthetic images poor training data. section 6.5
- Improve value by matching real recording conditions; rendering objects over augmented-reality recordings with accurately estimated lights can produce more realistic images. section 6.5
- Compare synthetic data with other bootstrap options, including similar-domain data, manual creation, dummy baselines, and third-party solutions. section 6.5
- Check whether generated data represents the target domain; a plausible synthetic source is not automatically appropriate for every problem. section 6.5
More realistic synthetic images
A straightforward renderer may produce unrealistic object images because lighting and shadows are imperfect. A stronger approach renders objects over augmented-reality recordings whose lights have been accurately estimated, making the images more valuable for model training. section 6.5
