Zappos 50k shoe image dataset: what UT-Zappos50K actually contains

One cataloged record answers this query exactly. UT-Zappos50K - 50,025 Shoe Images with Fine-Grained Attributes is the UT Austin Vision Lab benchmark built from a circa 2013-2014 crawl of the U.S. Zappos.com catalog: 50,025 product photographs spanning four broad categories (Shoes, Boots, Sandals, Slippers) and their sub-categories, mostly 136x102 pixels, centered on a white background and shot in a consistent orientation.

Datadory scores it 9 out of 10 on its 0-10 rubric (field documentation, access reliability, freshness) - a mark shared by 534 of the 1,744 datasets we catalog. Among the seven primary footwear records, only three reach 9, and this is the only one of them that ships images rather than statistics.

1. ut-zap50k-images.zip (291 MB) - all 50,025 full-resolution JPGs; both image archives are marked ACADEMIC USE ONLY.

1. ut-zap50k-images-square.zip (137 MB) - the same 50,025 images square-cropped, which spares you a centering pass.

1. ut-zap50k-feats.zip (208 MB) - precomputed features per image: 960-dimensional GIST descriptors plus 30-dimensional LAB color histograms.

1. ut-zap50k-lexi.zip (203 MB) - the fine-grained lexicon v2.1 with ten attributes (comfort, casual, simple, sporty, colorful, durable, supportive, bold, sleek, open), each holding roughly 4,000 ordered pairs.

1. ut-zap50k-data.zip (~3-8.6 MB) - meta-data.csv, meta-data.mat, the binary expansion meta-data-bin.csv, image-path.mat, comparison matrices zappos-labels.mat / zappos-labels-fg.mat, zappos-fg-rationale.mat, train-test-splits.mat and demo_script.m. Datadory's verification downloaded and unzipped it in-session: 10 files totalling 86,606,779 bytes.

The metadata table documents nine columns, one row per image: CID (a ProductID-ColorID key such as 100627-72, where several colorways share one ProductID), Category, SubCategory, HeelHeight, Insole, Closure, Gender, Material and ToeStyle. Only Category and SubCategory are guaranteed present; sample oxford rows show semicolon-joined multiples like Leather;Padded;Removable insoles. Formats split two ways: CSVs open as plain text, while every label, split and feature .mat file needs MATLAB or scipy.io.loadmat. All nine columns were checked against the archive itself, so field definitions carry a verified confidence flag.

Three corpora cover shoe imagery in Datadory's pool of 15 records (7 primary plus 8 related), and they solve different problems. The trade-off is depth versus convenience versus currency:

Four task families run cleanly on this release, and each maps to a different part of the archive:

1. Category and sub-category classification - 50,025 rows labelled with guaranteed-present Category and SubCategory values give supervised training data at a scale Fashion-MNIST matches only in count, not resolution.

1. Attribute prediction - the nine-field SKU table turns HeelHeight, Insole, Closure, Gender, Material and ToeStyle into multi-label targets, including semicolon-separated multiples.

1. Relative-attribute learning - the crowd-sourced ordered pairs (roughly 4,000 per attribute across ten attributes) supervise ranking-style models that learn "more comfortable than" rather than yes/no labels, with one-sentence rationales attached per annotation.

1. Classical baselines without a GPU - the shipped 960-dimensional GIST descriptors and 30-dimensional LAB histograms let you fit linear models or gradient boosting before committing to deep networks.

Because the release is static, you can pin one snapshot in an experiment tracker and get bit-reproducible runs - a property shared by the 181 static records among the 1,744 datasets Datadory catalogs, and the reason fixed snapshots remain preferable for reproducibility even though they cannot track assortment drift.

Data Scientists & ML Engineers rank first in Datadory's tagging for this record (relevance 3 of 3): fine-grained visual models trained on 50,025 labeled images with pair supervision, plus the precomputed features for fast iteration. The persona-level workflow, including how this corpus ranks against the trade APIs in the same slice, sits on the data scientists footwear page.

Writing it up? "Fine-grained attributes" here means the sub-category visual properties defined in Datadory's fine-grained attributes glossary entry, operationalized by this dataset's lexicon v2.1.

Two sibling clusters cover the neighbouring questions: how to assemble a shoe product catalog JSON feed from schema.org Product JSON-LD when you need current listings instead of a frozen crawl, and where sneaker size level demand data surfaces size-by-size asks, bids and sales counts once your model needs market signal rather than images.