LabelBox — Cooler Refill
Detection & Dataset Validation

Bounding box annotation and dataset validation to train and improve automated stock detection models — enabling data-driven restocking decisions for retail operations at scale.

3
Stock Classifications
BBox
Annotation Method
↑ Accuracy
Model Performance
↓ Manual
Verification Effort

Project Overview

Workgen Solutions worked on a structured bounding box annotation and dataset validation project using LabelBox — a leading data labeling platform — to support the development and accuracy improvement of automated cooler refill detection models.

The objective was to build clean, accurately labeled image datasets that automation models could use to reliably identify product restocking levels inside retail coolers, reducing dependence on manual verification and enabling smarter, data-driven restocking workflows.

Each cooler image was precisely annotated with bounding boxes and classified into one of three stock-level categories, producing a high-quality, structured dataset ready for model training and validation.

Platform & Method

LabelBox Platform
Industry-standard data labeling and dataset management tool
Bounding Box Annotation
Precise rectangular region marking around cooler units and product areas
3-Class Classification
Full, Partial, and Empty stock levels with validated dataset output

Stock Level Classification System

Every annotated cooler image was assigned one of three clearly defined stock-level categories to ensure model training data was unambiguous and consistent.

Full
Fully Restocked

Cooler shelves are completely stocked with products. No visible gaps or empty slots. Products are standing upright and neatly arranged.

  • All shelf positions occupied
  • Products clearly visible and upright
  • No restocking action required
Full — No Action Needed
Partial
Partially Restocked

Cooler shelves are partially stocked. Visible gaps exist but some product remains. Restocking is needed but not urgent depending on level.

  • Some empty shelf positions present
  • Remaining product visible but insufficient
  • Restocking recommended soon
Partial — Restock Soon
Empty
No Stock Available

Cooler shelves are bare or near-empty. Immediate restocking is required. This classification triggers priority restocking workflows.

  • Shelves visibly bare or nearly empty
  • No usable product visible to customers
  • Immediate restocking required
Empty — Restock Immediately

Annotation & Validation Process

1

Image Ingestion & Dataset Setup

Cooler refill images were ingested into LabelBox and organized into structured batches for annotation. Each image was associated with metadata including store location and shift context.

2

Bounding Box Annotation

Each cooler unit and relevant product region within the image was precisely annotated using bounding boxes — defining the exact area the model should evaluate for stock level detection.

3

Stock Level Classification

Each annotated region was assigned one of three stock-level labels — Full, Partial, or Empty — based on clearly defined visual criteria, ensuring consistent labeling across the entire dataset.

4

Dataset Quality Validation

Completed annotations underwent structured quality validation to identify and correct mislabeled samples, poor bounding box alignment, or ambiguous classifications — ensuring model-ready dataset accuracy.

5

Dataset Export & Model Integration

Validated datasets were exported in structured formats compatible with the client’s automation pipeline, providing clean, labeled training data to improve cooler refill detection model accuracy.

Tools & Technology

LabelBox

Industry-standard data labeling platform used for bounding box annotation, dataset management, and quality review

Bounding Box Methodology

Precise rectangular region annotation to define cooler and product boundaries for automated model detection

Structured Dataset Export

Validated datasets exported in structured formats for direct integration into client automation model training pipelines

Results & Impact

Clean, validated annotation datasets directly improved the performance of automated cooler stock detection systems.

Improved Model Accuracy

Higher quality labeled datasets directly improved the accuracy of automated cooler refill stock detection models

Reduced Manual Verification

Automation models trained on clean data required significantly less manual review and correction in downstream operations

Data-Driven Restocking

Enabled smarter, classification-driven restocking decisions by providing reliable Full/Partial/Empty detection at scale

Enhanced Retail Execution Monitoring

Automation-powered monitoring improved visibility into real-time cooler stock levels across retail locations

Why This Matters

This project demonstrates Workgen Solutions' capability to deliver production-grade ML annotation datasets using industry tools like LabelBox. By applying structured bounding box annotation and rigorous dataset validation, our team created the clean, consistent labeled data that automation models need to perform reliably — directly enabling smarter retail execution at scale.