Retail store item detection isn’t easy due to factors such as differences in lighting conditions, camera angles for capturing images or video of items with your deployed model, occlusions, and products that may look similar, or come in different sizes.
Here are some considerations to help with creating a robust retail store item detection dataset and model:
- If you’re trying to distinguish between “milk”, “carrots”, and “apples” you probably don’t need that many examples. If you want to distinguish between “1% Anderson Ericson milk”, “2% Anderson Ericson milk”, “1% store-brand milk”, “2% store-brand milk”, “brand A baby carrots”, “brand B carrot strips”, “organic carrot strips”, etc you’ll need many more examples of each so that your model learns the subtle differences.
If your displays are all exactly the same with identical lighting conditions (eg. like a supermarket chain) then you’ll need less than if stuff is set up differently at every store (eg. detecting items in NYC bodegas).
More resources to help:
- A Guide for Model Production - Roboflow
- Labeling Guide: Object Detection - Roboflow
- Dataset Health Check: Improving Your Dataset - Roboflow
- Health Check: Filter by Class - Roboflow
- Retail Store Item Detection using YOLOv5 || Model Training: YOLOv5 Video Tutorial - Roboflow || Run YOLOv5 on Your Webcam - Roboflow
- Datasets: retail store | Roboflow Universe Search || grocery store | Roboflow Universe Search || https://universe.roboflow.com/search?q=SKU110k&t=metadata