Shelf audit tool: embeddings fail to tell sibling SKUs apart after YOLO detection

ryan7ait · reddit · 2026-09-28

A developer building a shelf audit tool reports that the identification stage is the hard part: YOLO detects and crops products fine, but embedding-based nearest-neighbor search against a reference gallery frequently confuses sibling SKUs — same brand and bottle, different flavor or size (1.25L vs 2L). Correct and wrong similarity scores overlap, so no threshold works.

DINOV2, SigLIP2 and OpenCLIP all failed the same way. Root causes: letterboxing crops to 224 destroys the tiny size text, and most SKUs only have a couple of real shelf photos as references. The author is asking whether to fine-tune the embedder on hard negatives, add OCR as a second check, or drop the global-embedding approach entirely.

Original post →

More from Apps

Apps channel →