Shelf Vision

Term definitionscoming soon

Definitions for this piece will land here. Inline terms in the body will link into this panel once they are filled in.

/** Fill in when the case-study rewrite lands. Used by ArticlePage Term definitions. */ export const glossary = [];

Finding products on a shelf is a detection problem for about ten minutes, and then it becomes a retrieval problem. Detection tells you something is there. The hard part is that forty of the somethings are the same brand in slightly different packaging, and a classifier with a fixed label set is obsolete the moment the store changes supplier.

The pipeline

1. Detect — YOLO proposes boxes. Recall matters far more than precision here; anything missed at this stage cannot be recovered later.
2. Retrieve — each crop is embedded and matched against a product gallery, so adding a new product means adding an image, not retraining a classifier.
3. Heuristic rerank — cheap structural signals: shelf row, neighbour agreement, box geometry.
4. Contextual rerank — products cluster. A crop between two of the same SKU is very likely a third.

Scored on COCO-format bounding boxes at mAP@0.5.

What I would defend

• Retrieval over classification. A gallery you can extend beats a label set you have to retrain, in a domain where the catalogue changes weekly.
• Rerank in two cheap passes instead of one expensive one. The heuristic pass removes most of the candidates the contextual pass would have to think about.

Result

PLACEHOLDER — the mAP figure and where the pipeline still breaks: reflective packaging, deep shelves, and anything at a sharp angle.