Key Insights: Reduced shelf analysis time by 30x and achieved over 90% classification accuracy.

About the Client

GU Trade helps mass consumption retail companies gain real-time visibility into point-of-sale activities so they can monitor KPIs effectively and make evidence-based decisions. To stay competitive, they needed to modernize their shelf monitoring process.

The Challenge

GU Trade’s staff periodically visited supermarkets to photograph shelves and check brand distribution. This manual process was slow, costly, and prone to human error. Images were often incomplete, unlabeled, or noisy, making accurate analysis even harder.

They needed an automated computer vision solution capable of determining a brand’s Share of Shelf (SOS) from photos — even when faced with low-quality, inconsistent data.

Marvik’s Approach

We designed and deployed a custom computer vision pipeline that:

• Cropped and labeled each product for precise counting.
  
  
• Used RetinaNet to detect items and avoid double counting.
  
  
• Applied OCR and CNNs for classification.
  
  
• Stitched partial shelf photos together for complete coverage.
  
  
• Augmented and cleaned data to handle challenges like blocked visibility, upside-down packaging, and inconsistent image quality.
  

The solution was deployed on Azure, leveraging Python, OpenCV, TensorFlow, Keras, and PyTorch.

The Results & Impact

• Processing time 30x faster than manual counting.
  
  
• Over 90% classification accuracy.
  
  
• 500+ shelves processed automatically.
  
  
• 50,000+ products detected with high confidence.
  

By replacing a manual, error-prone workflow with a scalable, high-accuracy system, GU Trade can now act on shelf data faster and with greater confidence — turning retail execution into a measurable competitive advantage.

Why This Matters

This project shows how AI, when tailored to a client’s reality and deployed with speed, can transform routine operations into high-impact, data-driven decision-making.