Shelf recognition looks like a solved problem until the taxonomy moves under it. A supplier renames a product line, a pack size changes, a seasonal variant appears, and a label set that was clean in January quietly stops describing what is on the shelf in June.
Ambiguity is the second failure. Two annotators looking at the same partially occluded pack will disagree, and if the guideline does not say which of them is right, the disagreement is baked into the training set as noise.
The fix is unglamorous: a versioned taxonomy, a written adjudication rule for every ambiguous case, and a reviewer loop that samples enough to catch drift before it reaches a release.