Ground Engaging Tools (GET) Bounding Box Annotation
Shaip delivered 40,000 bounding box annotations per month across 15 Ground Engaging Tool classes — bucket teeth, adapters, shrouds, lip shrouds — on backhoe-bucket imagery, distinguishing internal versus external faces under field-grade conditions of dirt, debris, motion blur, and low light, at a 99%+ accuracy gate in KITTI 1.0 format, across a 12-month wear-monitoring AI contract
Project Overview
As predictive maintenance and wear monitoring moved into heavy equipment AI, the client needed a domain-specific annotation pipeline capable of detecting and labeling Ground Engaging Tools (GETs) on backhoe buckets across noisy, dirty, real-world field imagery.
Shaip built the end-to-end annotation pipeline covering bounding box placement across 15 GET classes, golden-rule edge-case handling, two-level QA, and weekly KITTI 1.0 delivery — supporting a 12-month, 40K-image-per-month production cadence at a 99% accuracy gate.
Key Status
Images per Month
40,000
GET Classes
15
Accuracy Threshold
99%
Weekly Delivery
10,000
Challenges
- Annotating 40,000 images per month with 9–10 bounding boxes per image on average
- Handling 15 distinct GET classes with internal vs. external face differentiation
- Working with field-grade imagery — dirt, debris, motion blur, poor lighting, and difficult angles
- Maintaining 99%+ annotation accuracy under high-volume weekly delivery pressure
- Exporting in KITTI 1.0 format ready for direct model training integration
Solution
GET Schema & Class Differentiation
Shaip configured a 15-class GET ontology covering both internal and external bucket views. Annotators were trained to distinguish between the internal and external faces of each component — a domain-specific differentiation critical for downstream wear monitoring accuracy.
Golden Rules for Field Imagery
Annotators followed strict golden rules to maintain data integrity. No annotations were created where components were not clearly visible. GETs covered by more than 50% of debris were skipped. Blurry or dark components were excluded. Lip classes were only labeled when 100% of the object was visible. Zoom was used only for navigation, not for labeling decisions.
Edge-Case Training Library
Fourteen detailed case studies were documented to train annotators on common edge cases — partially occluded GETs, buckets in motion, dark backgrounds, and boundary ambiguity between adjacent components. This library made the annotation guidelines one of the most comprehensive in the heavy equipment AI domain.
Two-Level QA & 99% Accuracy Gate
After Level 1 annotation, all images passed to a dedicated QC team for structured review. Minor errors were corrected with annotator feedback; major errors triggered rework reassignment. Datasets were approved only when annotation accuracy reached 99% or above. All QC outcomes were logged in a tracking sheet for full transparency.
Weekly Delivery in KITTI 1.0 Format
Weekly delivery targets were strictly maintained at 10,000 images, all exported in KITTI 1.0 format and ready for direct model training integration. The 12-month contract maintained this cadence with no slippage.
Project Scope
| Dataset Type | Monthly Volume | Classes | Avg Boxes / Image | Format | Contract |
|---|---|---|---|---|---|
| GET bounding box annotation | 40,000 images | 15 GET classes | 9–10 boxes | KITTI 1.0 | 12 months |
Outcomes
- Established a 12-month, 40K-image-per-month pipeline for heavy equipment wear detection AI
- Standardized 15 GET classes with internal/external face differentiation
- Maintained a 99%+ accuracy gate across every batch under weekly delivery pressure
- Built a 14 case study edge-case library for annotator training and onboarding
- Delivered KITTI 1.0 format outputs ready for direct model training integration
Overall, Shaip helped transform a field-grade GET detection requirement into a structured, production-ready annotation pipeline — one capable of supporting wear monitoring, predictive maintenance, and construction equipment AI initiatives with consistency, traceability, and field-realistic accuracy.
Shaip handled the messiest imagery we had — dirt, debris, motion blur, the works — and still delivered 99%+ accuracy on a 12-month, weekly-delivery cadence. Their KITTI 1.0 outputs went straight into model training.
– Product Manager, Heavy Equipment AI
