ViT-B/32
model
ViT-B/32 is the vision model we use to turn each product photo into a fingerprint. It is a middle-ground size: accurate enough for 20,000 products, light enough to run cheaply. We picked it over larger models for the best speed-to-accuracy on our hardware.
For engineers: a Vision Transformer, Base size, 32-pixel patches, fine-tuned with a metric-learning loss (ArcFace).