Synthetic data, computer vision, and the work behind them.
Engineering writeups, dataset experiments, and field notes on building production-grade vision AI with synthetic data.
All posts
3 posts · sorted by most recent
- Engineering
Why Defect Detection Models Fail on the Defects That Matter
Industrial inspection models break on rare defect types, not common ones. How a 1,000-image synthetic dataset across cables, transistors, nuts, and screws reached 0.90 mAP trained only on synthetic data.
- ADAS
The July 2026 Driver-Monitoring Mandate Has a Data Problem
From July 2026 every new EU vehicle needs a camera watching the driver. The privacy rules that make those systems lawful also make the training data illegal to collect. Synthetic data is the way out.
- Computer Vision
You Cannot Collect a Theft Dataset Without Filming the Theft
Real shoplifting footage is rare, privacy-loaded, and unevenly captured. A 571-image synthetic dataset across three theft behaviors and multiple camera views shows how to build coverage you cannot lawfully record.
