A prototype, with uncertainty in view.Found: presence ≥ 90%. Not found: ≤ 20%. Between them: unknown. A failed request is also unknown, never a negative match. Explicit wrong-brand reviews and rejected brand checks override a positive model score to unknown; their returned crops remain available as rejected evidence.
Boxes below 50% confidence, without reported confidence, or touching a crop edge are dashed and flagged for review. This is a demo rule, not a validated accuracy threshold. One representative container is located per product; this does not count facings.
Presence checks the full photo, then overlapping shelf crops with up to six independent requests in parallel. Dense crops run in stable batches of six. A crop scoring at least 97% stops scheduling later batches; every started request finishes and is counted. A coarse positive alone cannot stop the crop search. The scan runs at realtime speed without presentation pauses. Uncertain candidates get closer views when their best dense-crop score is above 20%; weaker candidates remain unknown. Close-ups include separate brand support; both scores must clear the same 90% cutoff. Box runs only after positive presence and can retry on a closer context view. After localization, a close-view brand guard withholds a candidate if brand support is ≤50% or unavailable; support below 90% flags the location for review. This extra guard never lowers the 90% match cutoff. Highest-of-many scores are not calibrated accuracy guarantees.
Jev Box documentation ↗