01
Your photo library already contains business knowledge
Teams sort product, inventory, and catalog photos every day. The useful labels are the ones your business already uses—not a generic list of objects. StayCharted AMT learns from pictures paired with those approved categories.
| Illustrative picture | Your category |
|---|---|
| Studio photo of a backpack | Backpacks |
| Front view of a wristwatch | Watches |
| Pair of running shoes | Sports Shoes |
02
A repeatable workflow for new pictures
- Collect labeled product photos
- Check examples and train
- Evaluate set-aside pictures
- Submit new photos or links
- Review categories before catalog updates
Process new images through file uploads or call the model from your application. Your application decides how predictions update a catalog, inventory tool, or review queue. A developer can connect the predictions to your catalog through the API.
03
Choose how your image model learns
With the Vision add-on, an AI Image Classifier Model uses ready-made vision AI to read pictures and a private classifier to learn your categories. The ready-made AI does not train on your uploaded pictures.
A Dedicated AI Image Model, on Business with Vision, trains the vision AI itself on your pictures. Training takes longer. Compare both approaches on the same set-aside pictures to choose for your task; dedicated training does not guarantee better accuracy.
04
Prepare pictures your model can learn from
Upload a ZIP with one folder per category, or a spreadsheet of public HTTPS picture links and category labels. Use consistent labels and examples that represent the lighting, angles, backgrounds, and quality of future pictures.
StayCharted AMT checks for category mismatches, conflicting labels, near-duplicates, and unreadable links. Review those findings before training. Hidden image metadata is removed on upload; this does not hide faces, personal information, or text visible in a picture.
05
Test the categories that matter to your catalog
Measure accuracy on pictures the model did not train on. Look at errors by category, including products that look similar. Keep near-duplicate shots of the same item from making evaluation appear easier than future work.
Use confidence scores to prioritize review. Confident results can still be wrong. Recheck performance when suppliers, photography styles, or your category definitions change.
06
Common questions about image classification
Does it identify an exact SKU? This workflow predicts your category labels. It does not establish that two pictures show the same individual item or perform exact product matching.
Can it read labels or documents in the picture? Vision classifies what a picture shows. It is not OCR. Use extracted text with a text model when the words determine the category.
Do I need a developer? You can prepare pictures, train, and process files without writing training code. Connecting predictions to a catalog application through an API may require a developer.
Will it work on my photos? Results depend on your examples and category distinctions. Test representative, unseen photos before relying on predictions.