Computer Vision
Retail shelf intelligence and inventory monitoring
Upload a shelf photo and get product counts, empty zone detection, and stock-level alerts — no manual audit needed.
Problem
Field teams armed with clipboards cannot scale across hundreds of SKUs and dozens of aisles. Existing inventory systems rely on POS data that only reflects what was sold, not what is missing right now. The challenge is a vision system that handles real-world shelf variation — different lighting conditions, densely packed products, rotation across store formats, and delivers actionable alerts fast enough to matter.
System demo
Inference running on SKU-110K public dataset · YOLOv8m pretrained checkpoint · Annotations rendered with OpenCV
Architecture
Shelf image upload
Region detection
Product / facing count
Empty zone flagging
Alert generation
Dashboard export
MetricsPrototype benchmarks measured on SKU-110K public dataset — to be replaced with production results
mAP@50 · product detection
84.7%
mAP@50 · empty zone detection
89.1%
Precision
88.3%
Recall
85.6%
Inference (GPU)
42ms
False stockout rate
6.2%
Tech stack
Edge deployment — we've shipped this on hardware
On-device inference keeps raw footage on-premises, eliminates cloud latency, and works reliably in stores with poor connectivity. We are experienced deploying vision models on edge hardware across a range of devices and can assess, recommend, and configure the right platform for your throughput, power budget, and environment.
Production considerations
- Camera placement — fixed ceiling or aisle-end mount recommended over handheld; consistency directly impacts model accuracy
- Lighting variation — augmentation pipeline handles brightness and contrast shifts across store formats and times of day
- SKU rotation — new product introductions degrade accuracy over time; a retraining trigger and workflow is built into every delivery
- Alert threshold tuning — high-velocity SKUs (beverages, dairy) need tighter thresholds than slow movers; configurable per category
- Dashboard integration — alert output can feed your existing store management system, POS platform, or BI tool via webhook or REST API; Streamlit is the default for standalone deployments
- Privacy — on-device inference means raw footage never leaves the store; only structured alert data is transmitted
Explore this project
Live demo and source code links will be added as they become available.