Fashion-MNIST - Zalando Research Apparel Image Dataset
Datadory delivers apparel, accessories and luxury goods data covering Fashion-MNIST: 70,000 labeled 28x28 grayscale product images across ten classes - T-shirt/top through ankle boot - split 60,000 train and 10,000 test as the drop-in replacement for the original MNIST benchmark. Get a sample of this dataset, delivered daily, weekly, or hourly.
Fashion-MNIST - Zalando Research | Industry: Apparel, Accessories & Luxury Goods
Fashion-MNIST is Zalando Research's apparel image benchmark: 70,000 examples of merchandise imagery reduced to 28x28 grayscale, each carrying a label from ten classes drawn directly from Zalando's own article inventory. It was built in 2017 for one stated reason - the original MNIST digit set had become too easy and too overused to judge modern computer vision - and it deliberately keeps MNIST's exact image size and train/test structure, so any pipeline written for digits runs on clothing unchanged. Ten classes, two fields per record, one frozen snapshot: it remains the fastest way to prove a vision pipeline works before committing GPU budget to bigger photographic corpora.
Sample rows
Five records straight from the shipped sample, shown as label and class pairs. Every one of the 70,000 examples carries exactly this shape - a pixel matrix plus one integer.
label=0 class=T-shirt/top
label=5 class=Sandal
label=7 class=Sneaker
label=8 class=Bag
label=9 class=Ankle boot
The full corpus holds 60,000 training records and 10,000 test records in the same form, so a stratified sample slice preserves all ten classes in their true proportions.
Field dictionary
Two fields carry the entire dataset. Both are documented and verified against the shipped files - there is no hidden third column to discover later.
| Field | Type | Definition | Example |
|---|---|---|---|
| image | integer matrix | 28x28 grayscale pixel values, 0-255, stored as IDX3-ubyte | [0, 0, 0, ..., 78] |
| label | enum 0-9 | integer class index mapped to ten apparel and accessory classes: 0 T-shirt/top, 1 Trouser, 2 Pullover, 3 Dress, 4 Coat, 5 Sandal, 6 Shirt, 7 Sneaker, 8 Bag, 9 Ankle boot | 9 |
Additional fields on request: decoded per-class image sets rendered as PNG, one-hot encoded label vectors, and explicit train/test split markers are not part of the raw distribution but can be added to your delivery - ask for them when requesting a sample.
Coverage
- Geographic: imagery originates from Zalando's product catalog; individual records carry no geographic attributes
- Temporal: a single snapshot released August 2017 - the contents are fixed, which is precisely what keeps benchmark scores comparable across years
- Granularity: one classification record per image - 60,000 train and 10,000 test examples, roughly 30 MB compressed (26 MB train images plus 4.3 MB test images plus the label files)
Delivery
API, files, or your warehouse. Daily, weekly, or hourly.
For a frozen corpus like this one the sensible default is a one-time file drop into your warehouse, refreshed on your schedule rather than anyone else's. Ask for the full 70,000 records or a stratified sample first - the format follows your decision, not the other way around.
Who uses Fashion-MNIST data
Four kinds of teams get real work out of this corpus, ranked by fit:
- Data scientists and ML engineers treat it as the drop-in MNIST-style benchmark for image models - validate a classification pipeline end to end on 70,000 labeled images before spending GPU time on larger catalogs. The data scientists use cases page expands the workflow.
- Developers and data-product builders reach for it when one loader call needs to replace an ingestion project; the two-field schema means a working demo in an afternoon. Detail lives on the developers builders use cases page.
- Journalists, academics and students cite the canonical 2017 paper by Han Xiao, Kashif Rasul and Roland Vollgraf (arXiv 1708.07747) as a standard object alongside reported numbers. See the journalists academics use cases page.
- E-commerce operators prototype automatic product categorization against a clean baseline before pointing models at messier live catalogs - context on the e-commerce operators use cases page.
The honest exclusion list is equally instructive: sales and growth teams, investors and quant researchers, market researchers and competitive-intel teams score zero here, because a frozen 2017 snapshot of product photos answers none of the questions those personas ask of apparel data.
Questions buyers ask
What are the ten classes in Fashion-MNIST?
The labels run 0 T-shirt/top, 1 Trouser, 2 Pullover, 3 Dress, 4 Coat, 5 Sandal, 6 Shirt, 7 Sneaker, 8 Bag and 9 Ankle boot - garments and accessories taken from Zalando's article inventory. Each integer pairs with one 28x28 grayscale image, giving every class thousands of training examples in the 60,000-record training half.
How large is the Fashion-MNIST corpus?
Roughly 30 MB compressed: 26 MB of training images, 4.3 MB of test images, plus the two label files - 70,000 images across the 60,000/10,000 split. Small enough to iterate on a laptop, large enough that class balance and model capacity actually matter.
What accuracy can models reach on Fashion-MNIST?
The project's results ledger tracks community submissions from roughly 0.883 with a multilayer perceptron up to 0.967 with a WRN40-4 network, against a crowd-sourced human baseline of 0.835. An automated scikit-learn harness reproduces scores for 129 non-deep-learning classifiers, so the gap between classical and deep methods stays directly measurable.
Why was Fashion-MNIST created?
Zalando Research's stated motivation was that the original MNIST digit set had become too easy and too overused to benchmark modern computer vision honestly. Fashion-MNIST swaps handwritten digits for real merchandising imagery while keeping identical image dimensions and train/test structure, so existing digit pipelines run unchanged on the harder task.
Does Fashion-MNIST change over time?
No - the contents are a single snapshot released in August 2017 and have stayed byte-stable since. That permanence is the feature: accuracy recorded today remains comparable with numbers published years ago, unlike catalogs rebuilt from live retail feeds, whose contents drift seasonally. Teams needing live assortment signals pair this baseline with a commerce-side feed.
How does Fashion-MNIST compare with other apparel image datasets?
It trades scale for frictionless benchmarking. DeepFashion offers 800,000+ annotated images across 50 categories and 1,000 attributes; Fashionpedia packages 46,781 detection images carrying 342,182 bounding boxes; the Fashion Product Images Dataset pairs 44,000 real photographs with attribute tables. All three demand heavier pipelines than a 30 MB two-field corpus. The head-to-head lives at our comparison of Fashion-MNIST versus the 44k product images dataset.
See the rows before you pay anything.
Name this dataset and we send real records from it — scoped to the fields you asked for.