ML / Done
Wind Turbine Defect Dataset (100K VLM)
Overview
Co-led dataset construction for a vision-language model that spots defects on wind turbine blades: 100,000 real drone-inspection photos across 14 defect classes, from LEP erosion to laminate cracks. With one fellow lead, I ran 5 mentors and 35 annotators across Indonesia: writing inspection guidelines, recording tutorials, holding daily QC syncs, and bridging the company team with the workers on the ground.
Challenge
Keeping 40 annotators consistent across 14 visually similar defect classes, training mentors faster than the labeling queue grew, and holding daily throughput without letting bad labels through.
Result
Delivered the full 100,000-photo batch at 98.8% QC validity: a 14-class defect taxonomy verified by domain experts, plus an annotation pipeline the team could keep running.
Key Statistics
100,000 real drone inspections
Photos
40 across Indonesia
Annotators
14
Defect Classes
98.8%
QC Validity
Technologies
Annotation
- CVAT
Data
- Drone Imagery
Model Target
- Vision Language Models
Gallery
Weekly Sync With Company Team
Weekly alignment between the Korean company team and the Indonesian annotation crew.
QC Review Session
Reviewing edge-case labels and throughput with mentors and the company team.
