Transforming Business with AI

Labeled 500K frames in 8 weeks achieving 92.4% mAP model accuracy and reduced MVP timeline by 2 months.

Challenge

The client needed to train computer vision models for object detection in variable lighting and cluttered warehouse conditions. They had over 500,000 frames of unlabeled footage, but lacked the internal bandwidth and domain expertise to annotate data effectively.

Solution

• 1
  Developed a detailed annotation taxonomy specific to warehouse inventory (e.g., pallets, crates, humans, forklifts)
• 2
  Set up a cloud-based labeling platform with integrated QA workflows
• 3
  Trained a team of domain-knowledgeable annotators on labeling protocols
• 4
  Built an automated pre-labeling script to speed up human QA

Execution

• Labeled over 500,000 frames in 8 weeks with >97% consistency
• Weekly reviews with the client's ML team ensured alignment on edge cases
• Delivered balanced datasets and metadata ready for YOLOv8 and EfficientDet model training

Results

Key Metrics