DeepFashion Dataset (CUHK MMLab)

It is the clothing database behind a decade of fashion computer vision, assembled by the Multimedia Laboratory at The Chinese University of Hong Kong and released with the CVPR 2016 paper "DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations." More than 800,000 images are organised into four core benchmarks: Category and Attribute Prediction (289,222 images), In-shop Clothes Retrieval (52,712 images covering 7,982 items with roughly 200,000 cross-pose and cross-scale pairs), Consumer-to-shop Clothes Retrieval, and Fashion Landmark Detection.

The release kept growing after the paper: a Fashion Image Synthesis benchmark arrived in October 2017, the MMFashion open-source toolbox followed in 2019, parsing masks, dense pose and fine-grained attributes were added in 2020, and DeepFashion-MultiModal landed in June 2022 with human-parsing labels, keypoints and textual descriptions. In Datadory's catalog it scores 9 out of 10 — a band shared by 534 of the 1,744 cataloged datasets.

Seven annotation families, detailed in the dictionary above. Two design choices give the corpus its staying power. First, both label spaces are closed sets — 50 categories, 1,000 attributes — so the label space is fully enumerable before you write any code, and confusion between predicted classes maps onto a real merchandising taxonomy rather than arbitrary clusters. Second, every image carries its evaluation_partition, which means any accuracy figure you compute is comparable against a decade of published results on the same splits.

The pair structure is the part most catalogues simply do not have. Over 300,000 images are linked cross-pose or cross-domain to other photographs of the same garment — a blouse photographed flat in a shop and again worn on a street share a key. That linkage is what makes retrieval, virtual try-on research and pose-invariant matching trainable here and not elsewhere.

It is static by design. Benchmark versions date to 2016–2017, with extensions layered through June 2022; nothing inside the corpus updates on its own. Static snapshots are common in machine-learning corpora: 181 of the 1,744 datasets in Datadory's catalog never update, and research releases cluster heavily in that group.

Plan around a frozen reference set rather than an incremental feed, and say so in your sample request if recency matters — Datadory pairs it with living complements from the same industry brief, such as Lyst's catalogue spanning more than 27,000 brands or H&M's transaction release carrying 31,788,324 purchases across 105,542 articles.

Five of Datadory's eight personas clear the relevance bar for this record, ranked here by fit:

1. Data Scientists & ML Engineers (relevance 3/3) — 800,000 annotated images with landmarks and cross-pose pairs, plus published splits to benchmark against; see data scientists use cases.
2. E-commerce Operators (relevance 2/3) — the taxonomy behind virtual try-on and automated tagging R&D, before committing to a labelling programme of your own.
3. Developers & Data-Product Builders (relevance 2/3) — well-defined partitions and documented annotation formats make it unusually straightforward to wire into a CV engineering pipeline; more on developers builders use cases.
4. Market Researchers & Consultants (relevance 1/3) — the category and attribute scheme doubles as an industry taxonomy of how garments are described and distinguished.
5. Journalists, Academics & Students (relevance 2/3) — one of the most heavily cited academic benchmarks in fashion vision, useful as a provenance-clean reference point in teaching and reporting.

Persona-level detail lives on our competitive intel product teams use cases page and the full apparel, accessories & luxury goods hub.

On scale and annotation depth it has no peer in the industry slice. Fashionpedia brings 46,781 images with 342,182 bounding boxes and a 294-attribute ontology — richer per-image segmentation, far fewer images. The Kaggle and Hugging Face copies of the Myntra catalogue offer 44,072 labelled SKUs that load as one flat table, but stop at coarse merchandising attributes with no geometry. Fashion-MNIST remains the 70,000-image grayscale baseline you graduate away from.

Where DeepFashion demands more of you is shape: it arrives as four benchmark-specific sets of image archives plus separate annotation lists rather than a single tidy table, and it is a frozen snapshot. Teams that want tabular convenience start with the Myntra copies and move up once the problem genuinely needs landmarks, dense annotations or cross-pose supervision.

Datadory scores it 9 out of 10 on its rubric of field documentation, access reliability and freshness — the same band as 534 of the 1,744 cataloged datasets, comfortably above the catalog-wide mean of 7.81. Field definitions are marked verified: they were cross-checked against the project's annotation-file documentation during research. Sample data exists and is quotable — img/Sheer_Pleated-Front_Blouse/img_00000001.jpg, labelled Blouse with a-line; sheer attributes, assigned to the training partition.

Two things keep it from a perfect score. The corpus is static, so freshness depends entirely on how you pair it with living sources; and the imagery was collected from the open web without geographic stratification, so coverage claims should stay at the global-consumer-photography level rather than market specific.