Using Unity to Generate Synthetic data and Accelerate Computer Vision Training
summary
Unity generates annotated training data with the Perception SDK / perception package, using domain randomization to vary lighting, backgrounds, occlusion and camera pose. A faster R‑CNN trained on a mix of 400,000 synthetic images plus a small real set (about 760 real examples) yields higher precision/recall than real data alone. Labelers produce 2D/3D boxes and segmentation automatically, dataset insights gives aggregate checks, and a customer aerospace project ships a validated 10,000 images dataset for continued generation.
AI-written summary · about
chapters (10)
- 0:00The data problem in AI
- 2:46Labeling and real data
- 6:34Synthetic data solution
- 7:53Workflow comparison
- 11:11Retail case study
- 14:47Aerospace application
- 17:20Unity perception package
- 18:21Demo and walkthrough
- 32:19Summary and roadmap
- 34:03Q&A session
related talks
43mBuilding Environments for Autonomous Vehicle Training
27mAI for social good
18mCustomizable Computer Vision Expands Data Access Without Compromising Privacy
30m🤯 No data? No problem! Synthetic data to the rescue
29mSynthetic Data in Vision AI: Theory and Practice
33mNext-generation tools for synthetic data and AI training
20mSynthetic Data: The new Super Power for AI
17mBuilding Training Data for Autonomous Driving