Mapping what every consumer audience actually wears

The formal name is respondent-driven sampling, and survey science reaches hidden populations with it: people no register has ever listed, who can only be found through each other. Our version starts from a small set of confirmed New Yorkers, reads the comment sections under their posts to find the next wave, and repeats. Every candidate is scored on two surfaces at once, how the account writes and what its thumbnails look like, and is admitted if either clears the bar. Each account we keep carries its own paper trail: which wave it arrived in, which account referred it, how close it sat to the centre of the audience we were aiming at. New York took 876 accounts and 4,464 posts.

The literature also comes with estimators that correct for the bias chain referral introduces, because a referral chain over-samples the well connected. We have not finished implementing those, and that is the honest state of this layer. What the method already buys is a sample drawn through real social ties instead of through a ranking algorithm, which is a different and far more correctable starting error than search gives you. Search returns whoever performs the city best. This returns whoever is actually in it, plus a bias we can name and eventually correct for.

The detector is a YOLO model we fine-tuned ourselves, trained on a hand-labelled set plus a multi-teacher pipeline that generates labels at volume. On our own benchmark it scores F1 0.597, and a smaller two-teacher variant lifted shoe recall from 0.21 to 0.80, footwear being the category most general-purpose detectors quietly drop. Both beat Gemini flash-lite at the same task. A person pass runs first, so clothing is found on bodies rather than anywhere in the frame, and when a single garment comes back detected as several pieces we union the fragments rather than discard them.

The real constraint on this layer is not the model, it is pixels. 81% of our crops came from 300 by 400 thumbnails with a median short side of 94 pixels, and no detector recovers detail the platform never transmitted. We now re-fetch origin-resolution covers where they exist, which gives a median 3.6 times more short side to work with. Detection error is also the largest single source of bad nodes further down: when we started judging whole clusters instead of pairs, 23% held the wrong category, and 57% of those were entire clusters of the wrong object. That is this layer failing, not the clustering.

The first version of this layer had a quiet, fatal flaw: the vectors keyed on the photograph rather than the garment. A shirt and a pair of trousers cropped from the same photo scored 0.680 cosine similarity, while two different shirts from two different photos scored 0.643. The photo was beating the clothing. Fixing it needed no new model. Cut the crop at the raw detector box with no padding, subtract 0.3 of the frame's own embedding, and photo bias falls from 0.672 to 0.428 while pair accuracy rises to 0.959. We held Paris out of the derivation entirely to check we had not tuned to New York, and every measure improved there too.

Going greyscale and partitioning on garment type instead of colour then took item separation from 0.469 to 0.762, which made the expensive masking step we had planned redundant before we built it. Then we changed the ruler and the picture got worse: judged as whole clusters rather than as pairs, the shipped graph is 37.6% usable nodes and only 11.7% exactly one product. One further result is worth stating plainly. Requiring three near-duplicates before a garment becomes a node removes 18% of crops, and it removes them unevenly, with dresses retained at 56.8% against sunglasses at 95.5%. A dress nobody else in New York owns cannot find two twins. A plain white tee always can, so the map under-reports exactly the garments that make a wardrobe distinctive.

Nothing gets named off a single lookup. A brand classifier proposes candidates, the strongest go out to visual search, and whatever comes back is settled by a rank-weighted vote across all the returns rather than by the top result alone. A second vote of the same kind decides what the item actually is, its category. A separate gate then rules on whether the garment is nameable at all, because most clothing honestly is not: a plain tee has no product identity to find, and a system that always produces an answer is a system that invents them.

We ran a hard test on the standard objection to this layer, which is that studio product photography and 80-pixel social crops sit too far apart for any shared embedding space to bridge. That turns out to be false. A crop from the wild retrieves shop photography of its own garment at recall@1 between 0.87 and 0.90, against a chance rate near 0.01. But when we hand-audited the top hits one at a time, right-kind retrieval came out around 0.76 rather than the 0.90 the automated score implied, because the search engine's own returns include junk that the metric had been counting as success. The mechanism is proven, the accuracy is not yet measured, and we would rather publish that sentence than the flattering number. Five matches in that audit were exactly right, down to an Air Jordan 4 Retro OG and a New Era 9Forty Yankees cap.