• Refactored and optimized team’s auto-labeling pipeline; Significant reduction in code for auto-labeling/label-smoothing algorithms used across flight recordings
• Modified auto-labeling stack to fuse existing label sources into single “best label”; Resulted in average performance error of ~0.03 IoU against human-sourced labels (best to date)
• Created tooling to quantify label performance (IoU, center-aligned IoU, etc), performance
ceilings, and correlations with in-flight telemetry across all datasets
