Most warehouse visual errors are not random. They are twins: same black charger brick, same brown carton, same generic bottle with a label you cannot read from three feet away.
If you train on random batches forever, you will keep getting 92% accuracy and the same 8% of SKUs will keep eating the day.
Let operators label the twins
The "not this item" button should store query image, wrong SKU, and right SKU when they pick the correct card. That triplet is a hard negative you would never have found sitting at a laptop.
I review those pairs weekly. If two SKUs collide more than a handful of times, they get extra floor photos, and I add a simple rule: if the matcher is torn between them, force a human pick and do not auto-advance.
Do not overfit the twins in isolation
It is tempting to fine-tune only on the failures. Then you break the easy SKUs. I keep a mix: confirmed dwells as positives, twin rejections as hard negatives, and a slice of easy barcodes as an anchor so the model does not forget the obvious.
