How visual dupe matching actually works

The naive approach to “do I own this already?” is to compare product titles. It fails immediately: a “belted wool trench” and a “longline tie-waist coat” are the same coat, and two items labelled “white shirt” frequently are not.

Continue turns every catalogued photo into a descriptor built from silhouette outline, hem and sleeve length ratios, dominant colour family in CIELAB, pattern frequency, estimated fabric weight and detected brand. A candidate item on a product page gets the same treatment, and we compare descriptors rather than words.

Brand ends up being the weakest signal in the stack — useful for resale pricing, nearly useless for “is this the same thing”.