AI-Powered Shelf Price Verification: Matching Label Prices to POS Server Prices (blog.roboflow.com)

Roboflow published a practical workflow that turns a shelf photo into an automated price audit: detect each shelf label, crop it, run OCR to extract product names and prices, query the store’s POS server via API to validate entries, and return an annotated image with green boxes for matches and red for mismatches. The pipeline is built in Roboflow Workflows and chains object detection, a detection-offset expansion, dynamic cropping, Gemini-based OCR/structured parsing, and a final visualization step; when a label isn’t detected you can retake a closer photo to improve results. Technically, the demo uses custom Python blocks to call Google’s Gemini (gemini-2.5-flash) via the Generative Language API (requests + JSON) because the native Gemini block lacked support for that model at the time. Gemini returns normalized [y1,x1,y2,x2] boxes which are converted to pixel coordinates and packaged as a Supervision Detections object for downstream blocks; confidences are defaulted (Gemini doesn’t supply them), so the workflow adds an offset to boxes to boost OCR reliability before cropping. Operational notes: you must supply Gemini and POS API keys, run a local Roboflow inference server for custom code, and watch Gemini rate limits/billing. The result is a fast, reproducible alternative to manual price entry that can reduce human error and scale routine retail audits.