The tool compares a photograph against published museum objects and suggests which restricted category, if any, it most resembles. It is not an authentication, a provenance finding or an import determination. Uncertain results are referred to a human expert.
1. Photograph the object
Or try a held-out test image (never seen by the reference index)
Test images are held-out objects from the same museums. The last row is out of scope: objects not on any designated list, which the tool should refuse.
What the index covers
Measured on held-out images
Numbers are computed by embed.py on images the index never saw. The referral threshold was set on a calibration half of the test images and scored on the other half. Categories with small reference sets are labeled; their numbers are less reliable.