Detecting windshield damage with synthetic data

Windshield damage accounts for 30% of all auto-insurance claims inside a global industry worth roughly $37–41 billion. Real-world data for this domain is scarce and difficult to label because transparent and reflective surfaces are notoriously hard to capture, annotate, and validate.

We author scenes in USD and programmatically vary illumination, weather, camera pose, lens model, shutter, sensor noise, glass curvature, lamination, and damage taxonomy. NVIDIA Replicator produces RGB, depth, normals, instance masks, and metadata for automatic labeling. NVIDIA Cosmos-Transfer acts as a physical AI foundation layer that augments variability in a consistent way, giving us scalable datasets covering specific environments and edge cases via natural language as an extra input.

We trained a semantic segmentation model on an Omniverse synthetically generated dataset, bridging the Sim2Real gap without real data. Iteration during development was accelerated by 1.4× thanks to quick SDG turnaround, and our reworked stage randomization logic on top of Replicator achieves ~40% faster generation while preserving quality with the RTX Real-Time 2.0 rendering engine.

The resulting model provides powerful reporting capabilities with a segmentation mask — measuring the damage and its position on the windshield, and issuing a fully “notarized” repair vs. replace decision that serves insurance companies as a validation comparable to the valuator output.