• High complexity of labels: Required multi-level annotations (bounding boxes, keypoints, behavior tagging) across diverse real-world retail scenes.
• Large dataset volume: Client needed >2.5M frames annotated within a short timeline.
• Fine-grained actions: Picking up, putting back, misplacement, holding, inspecting product, requiring trained annotators.
• Consistency & accuracy: Needed ≥85% IoU accuracy to ensure model reliability across 300+ stores.
• Limited internal resources: Client lacked sufficient annotation capacity to meet deployment deadlines.

‍

• Deployed a dedicated annotation team: 20 annotators, 2 QA leads, and 1 PM to manage end-to-end workflow.
• Provided multi-layer annotation:
  • Bounding boxes for hands, SKUs, baskets
  • Keypoint labeling for human posture
  • Action classification for customer behaviors
• Used advanced tooling: CVAT with custom plugins, YOLOv8 pre-labeling, frame tracking automation.
• Implemented a 3-step QC process: IoU-based checks, random audits, and weekly QA review to maintain 96%+ accuracy.
• Optimized speed & accuracy: Automated pre-labeling improved throughput by 32%, enabling early dataset delivery.