Glossary

fine-grained attributes

Fine-grained attributes are the sub-category visual properties — shape, material, heel height — that separate visually near-identical shoes. UT-Zappos50K operationalizes them over 50,025 white-background Zappos catalog images with an attribute lexicon, per-SKU metadata and crowd-sourced pair comparisons; its v2.1 lexicon extension shipped 2017-11-02.

What are fine-grained attributes?

They are the visual variables that operate below the category level: two boots are both boots, but toe shape, material and heel height decide which one a shopper clicks. Datasets built for this task must therefore label within narrow categories, not across them.

UT-Zappos50K — full name “UT-Zappos50K - 50,025 Shoe Images with Fine-Grained Attributes” — draws 50,025 catalog images from Zappos.com shot on white backgrounds, attaches eight metadata fields per SKU, and ships precomputed GIST and LAB features alongside crowd-sourced relative-attribute pair comparisons. The attribute lexicon was extended on 2017-11-02 as version 2.1, marking the boundary between the original label set and the enlarged one.

Why do fine-grained attributes matter when choosing a dataset?

Catalog search and recommendation live or die on sub-category distinctions. A model trained on coarse labels returns “sandals” when the shopper asked for a low-heeled leather sandal, because the attributes that encode those differences were never in the training data.

  • Background protocol affects transfer. All 50,025 images sit on white backgrounds, matching catalog conditions but demanding care on cluttered photos.
  • Pairwise labels beat bare class labels. Crowd-sourced relative-attribute comparisons teach the model “more pointed,” not just “pointed” — the mechanism behind relative-attributes.
  • Lexicon versions must be pinned. The v2.1 extension of 2017-11-02 changed the label space; mixing eras corrupts evaluation.

How do you evaluate fine-grained attributes in a data source?

  1. Verify scale and shooting protocol. 50,025 images on white backgrounds is the UT-Zappos50K baseline; deviations limit comparability.
  1. Inspect per-SKU metadata depth. Eight metadata fields per SKU should accompany each image.
  1. Check which features ship. GIST and LAB features are precomputed; confirm whether you need raw-pixel retraining or can start from them.
  1. Pin the lexicon version. Note whether labels follow the original set or the v2.1 extension released 2017-11-02.
  1. Match the loading format to your stack. Folder-per-label layouts load directly as an imagefolder-dataset in common training frameworks.

Entries adjacent to fine-grained attributes in this glossary:

Frequently asked questions

What does the UT-Zappos50K dataset contain?

50,025 catalog shoe images from Zappos.com on white backgrounds, with eight metadata fields per SKU, GIST/LAB features, and crowd-sourced relative-attribute pair comparisons.

When was the attribute lexicon extended?

Version 2.1 of the lexicon extension was released on 2017-11-02, enlarging the attribute label space beyond the original release.

Datasets containing this field

Datasets containing fine-grained attributes

6 datasets carry fine-grained attributes in the catalog. Open one, count the fields, judge for yourself.

Datasets

StockX - Sneaker Resale Marketplace (Prices, Sales History, Bids/Asks)

Retail Price · Last Sale · 52-Week High / 52-Week Low …+4 more

Industries

USITC DataWeb: U.S. Footwear Import, Export and Tariff Statistics

HTS Number / HTS10 & DESCRIPTION · Customs Value (CONS_CUSTOMS_VALUE) · First Unit of Quantity (CONS_FIR_UNIT_QUANT) …+5 more

Industries

UT-Zappos50K — 50,025 Shoe Images with Fine-Grained Attributes

CID · Category · SubCategory …+6 more

Footwear

World Footwear Yearbook 2026 (APICCAPS)

Production / Consumption / Exports / Imports - Quantity (million pairs) · Exports / Imports - Value (million USD) · Average Price (USD) …+6 more

Every listing shows the field dictionary, sample rows, and coverage before you commit. API, files, or your warehouse. Daily, weekly, or hourly.

Get sample rows