Counting Groceries with Computer Vision
summary
Automated stock counting runs on 16 camera rigs feeding a FastAPI service and an LLM proxy, logging predictions to Snowflake and sampling a small percent for human recounts; metrics include exact-count accuracy, MAPE, RMS error and sign bias; adding the supply-chain image as context and an accountability check (precision/recall on countability) raises accuracy sharply; experiments compare traditional detectors, Gemini 2.0 flash and larger models, and show that a fine-tuned 2.5 flash plus a few simple rules meets accuracy and cost targets in production.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 1:51Virtual supermarket overview
- 7:59Automated warehouse overview
- 9:43Grocery inventory tracking challenge
- 11:08High-quality image capture
- 12:37Labeled dataset for stock counting
- 13:41Model evaluation metrics
- 15:27LLM data split strategy
- 16:11Traditional CV vs multimodal LLMs
- 18:41MMU benchmark selection
- 20:47Initial model performance
- 21:52Image quality accountability checks
- 23:37Rules for accuracy improvement
- 24:42Performance versus cost balance
- 26:28Model reliability in production
- 27:53FastAPI deployment
- 28:57Key learnings
- 30:04Closing remarks
related talks
34mRetail tracking edge2cloud system
47mUnlocking Data-Driven Apps with TensorFlow
32mKeynote: Everywhere commerce
25mProduct Management and AI
42mProduct Recognition in Retail/E-Commerce with Azure AI
40mMachines are Learning
58mPlaying the game with Azure Cognitive Services
36mApps That See 👀 Practical AI Vision for Developers