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

UT-Zappos50K — 50,025 shoe images with fine-grained attributes, cataloged by the UT Austin Vision Lab: 50,025 catalog photographs of Zappos.com merchandise, mostly 136x102 pixels, centered on a white background and shot in one consistent orientation. Four broad categories — Shoes, Boots, Sandals, Slippers — expand into their Zappos sub-categories (Oxfords, Loafers, Sneakers-and-Athletic-Shoes, Ankle Boots and the rest), so the corpus doubles as a labeled taxonomy exercise and an image corpus at the same time.

Three layers ship together, which is what makes this record unusual among footwear corpora. The attribute table carries eight fields per SKU, from heel height to toe style. The feature layer holds a 960-dimensional GIST descriptor plus a 30-dimensional LAB color histogram for every image — precomputed, so a baseline classifier runs before any GPU time is spent. And the supervision layer holds crowd-sourced relative-attribute pairs: two shoes compared head-to-head by five human annotators each, with a one-sentence rationale per judgment. Most product-image datasets stop at labels; this one records why humans preferred one shoe over another.

Get a sample of this dataset.

Three consecutive rows from the attribute table, exactly as they arrive in an extract — one row per shoe image, keyed by CID:

CID             Category   SubCategory   HeelHeight          Insole                     Closure    Gender   Material   ToeStyle
100627-72       Shoes      Oxfords                           Leather                    Lace up    Men      Leather    Capped Toe;Round Toe
100627-255      Shoes      Oxfords                           Leather                    Lace up    Men      Leather    Capped Toe;Round Toe
100657-72       Shoes      Oxfords                           Leather;Padded;Removable   Lace up    Men      Leather    Capped Toe;Round Toe

Read the keys, not just the values. The first two rows share ProductID 100627 and differ only in ColorID — 72 versus 255 — which is how the same Oxford appears twice as distinct images, once per colorway. Any join against your own catalog should key on the full CID, not the product prefix, or you will silently collapse colorways. Empty cells are honest signal: HeelHeight is blank here because these are men's flat Oxfords, and the field only populates where Zappos merchandisers recorded it.

Every one of the 50,025 rows carries this exact shape, and each CID resolves to its photograph plus its GIST vector, LAB histogram and pairwise judgments in the sibling files.

Eight documented attribute fields plus the CID key — nine columns carrying every row, all verified against the shipped files rather than inferred from documentation. Only two are guaranteed present on every record (the category pair); the remaining six populate where Zappos merchandisers recorded them, which is itself analytically useful: a populated Insole or ToeStyle marks a listing detailed enough that someone wrote construction notes.

Beyond the table sit three derived layers, each keyed back to the same CID. A 960-dimensional GIST descriptor and a 30-dimensional LAB color histogram per image, ready for nearest-neighbor and baseline classification work without touching a pixel. A ten-attribute lexicon — comfort, casual, simple, sporty, colorful, durable, supportive, bold, sleek, open — each attribute holding roughly 4,000 ordered pairs of images ranked by humans. And the relative-attribute supervision: four attributes (open, pointy at the toe, sporty, comfort), each pair judged by five workers, yielding 6,751 ordered pairs plus 4,612 equal pairs in the first slice and 4,334 fine-grained pairs in the second, with a one-sentence rationale attached to each annotation.

The canonical record stops at nine columns plus the derived layers above. Extracts can extend it mechanically, and anything below ships folded under additional fields on request rather than pretending to be core:

• Square-crop image set alongside the full-frame photographs, for pipelines that assume square inputs.
• Binary-expanded attribute table — the one-hot rendering of every multi-value cell, ready for model consumption without a parsing step.
• Predefined train/test split assignments, shipped as their own file so evaluation harnesses reproduce published baselines exactly.
• Typed decodings of the matrix-format supervision — comparison matrices, rationales and splits rendered as plain JSON or Parquet, so nothing in your pipeline needs a scientific-computing reader.

• Geographic — not geographic, and worth saying plainly. The corpus is the United States Zappos.com catalog assortment at collection time, so merchandising conventions are consistent end to end: American sizing language, marketplace-style material naming, a gender axis built for US retail. No records carry location attributes, because none of the questions this corpus answers need them.
• Temporal — static by design. Collection ran circa 2013–2014, and the fine-grained lexicon extension (v2.1) landed in November 2017. Nothing updates; that frozen quality is precisely what keeps benchmark scores comparable across years and papers.
• Granularity — one row per shoe image, 50,025 rows keyed by ProductID-ColorID, never aggregated. Attribute pairs live one level finer still, at image-pair level, so supervision arrives at the granularity the learning problem actually uses.

Scale check for planning purposes: roughly 291 MB of full-resolution photography, 208 MB of precomputed features, 203 MB of lexicon, about 3 MB of attribute table — small enough to hold in memory whole, large enough that class balance and annotation noise genuinely matter.

API, files, or your warehouse. Daily, weekly, or hourly.

For a frozen corpus like this one the sensible default is a one-time drop into your warehouse, refreshed on your schedule rather than anyone else's. The nine-column table arrives typed either way; images travel object-addressed or inline, your storage layer decides; and the feature matrices can land as native arrays or as plain JSON if nothing in your stack reads matrix formats. A sample ships first — rows, field dictionary, and a slice of the image set confirmed against your use case before any commitment.

Five kinds of teams get real work out of this corpus, ranked by fit:

1. Data scientists and ML engineers train fine-grained classification and retrieval models on 50,025 labeled images with human preference pairs already attached — supervision most catalogs make you collect yourself. Precomputed features mean a baseline exists before the first training run. Patterns like these sit on the data scientists use cases page.
2. E-commerce operators prototype visual search, similar-item rails, and automatic categorization against a real retailer's taxonomy — four categories expanding into dozens of Zappos sub-categories give a realistic label space, not a toy one. Context on the e-commerce operators use cases page.
3. Developers and data-product builders wire image-aware ingestion paths against a corpus small enough to move but structured enough to matter — one CID keying images, tabular attributes, vectors and pairwise judgments makes it a complete integration rehearsal. Detail lives on the developers builders use cases page.
4. Merchandising and assortment analysts read the attribute distribution itself: which closures dominate which sub-categories, how heel height spreads across genders, where material language clusters — a snapshot of how one major US retailer described its own inventory mid-decade.
5. Journalists and academics cite the canonical papers behind the corpus (Yu & Grauman's CVPR 2014 and ICCV 2017 work) as a standard object, and use the per-pair written rationales as a rare source of human explanations for visual preference. See the journalists academics use cases page.

The honest exclusion list: price and resale teams, demand forecasters, and competitive-intel groups score zero here — the corpus carries no prices, no sales history, and no timestamps beyond the crawl window. For market-side signal, the StockX resale feed and the USITC trade series in the same industry rail are the right neighbors, not this one.

Four things, all solved by taking it through Datadory instead. The supervision ships in matrix-file formats — comparison matrices, rationales and splits arrive as scientific-computing binaries that need a specialist reader before anything else can touch them; we decode them once and deliver plain rows. Multi-value cells hide inside single strings — Insole reads Leather;Padded;Removable and ToeStyle reads Capped Toe;Round Toe, so naive parsers treat construction details as opaque text; our extract splits them into typed lists while keeping the original string alongside. Field presence is uneven by design — only the category pair is guaranteed, and treating optional blanks as missing-data errors misclassifies thousands of legitimate flat shoes; the sample documents presence rates per field so joins behave predictably.

Fourth, the key structure punishes shortcuts: several colorways share one ProductID, so any aggregation keyed on the prefix silently deduplicates distinct products. Every delivered row carries the full CID, and the sample includes the parent-child mapping between products and their colorway images.

Cards worth reading next: the Zappos structured footwear catalog, whose ~50,000 SKU tabular attributes join naturally against these images on product identity; the StockX sneaker resale feed for price and sales-history context on the sneaker side of the industry; and the Hugging Face sneaker image record as a minimal smoke-test fixture when a one-row payload is the point. The industry hub collects the full rail for Footwear, and the best-of list ranks where each source lands.