Retail Shelf Recognition Dashboard POC

This proof-of-concept focuses on end-to-end recognition of every SKU on a retail shelf. From a single photo—whether it comes from a shopper’s phone or a fixed in-store camera—the system must detect each item, identify its exact product name and size, map its position on the shelf, and flag any empty facings.

The flow I expect:
1. Images are dropped into a watched folder or uploaded through a simple web page.
2. A lightweight computer-vision model (YOLOv8, Detectron2 or a comparable PyTorch/TensorFlow solution) parses the image and returns bounding boxes with the recognised SKU label and size.
3. Results feed straight into a browser-based dashboard (Streamlit, Dash, or similar) that charts per-SKU counts, visualises shelf positions, and highlights empty spaces in a contrasting colour. Metrics refresh in near-real-time so I can scroll through multiple images and instantly spot gaps.

Core deliverables
• Trained model (with weights) that correctly tags each product in our sample set by name and pack size.
• A small, self-contained dashboard project folder ready to run with one command.
• README covering setup, how to retrain on new images, and how empty facings are calculated.
• Demo video or live session proving it works on a mixed batch of phone snapshots and fixed-camera shots.

Acceptance criteria
• ≥90 % precision and recall across the supplied test images.
• Dashboard displays three data points per SKU: count, shelf coordinates, and empty-facing flag.
• Works on both portrait and landscape images without manual rotation.

Once the POC meets these markers I will green-light the next phase—automatic alerts and multi-store aggregation.