Fashion is the most data-hungry vertical in ecommerce AI. The category tree alone can run to 10,000 leaf nodes across a major retailer. Building effective models requires understanding not just what data to collect, but how to structure it.
The Fashion Category Hierarchy
A scalable taxonomy typically has three levels:
Level 1: Clothing → Level 2: Tops → Level 3: T-shirts / Shirts / Blouses
Level 1: Accessories → Level 2: Bags → Level 3: Tote bags / Clutches / BackpacksTrain at Level 2 for most classification tasks unless you have 1,000 + images per Level 3 node.
Fine-Grained Attribute Labelling
For fashion specifically, attributes matter more than for most other categories. Style recommendations, size guidance, and visual search all depend on accurate attribute tags. High-priority attributes per category:
- Tops: neckline, sleeve length, fit, pattern, colour
- Bottoms: cut, rise, fabric, length
- Dresses: silhouette, length, occasion, fabric weight
- Footwear: heel type, toe box, closure, material
Each attribute dimension in ImageHub's retail AI dataset maps to an ISO-standard controlled vocabulary to ensure consistency across sources.
Model On-Figure vs. Flat-Lay vs. Cut-Out
Fashion datasets contain three distinct image types that behave differently in training:
- On-figure (model wears garment) — most common retailer image, contains body pose as confound
- Flat-lay (garment on surface) — neutral background, useful for texture and detail extraction
- Cut-out (white background, no model) — common in marketplaces, clean but loses drape and fit signals
Always record image type as a metadata column. Models trained on a mixture without this flag learn spurious correlations between background and category.
Scale Requirements by Task
| Task | Minimum images per class |
|---|---|
| Level-2 category classification | 500 |
| Attribute extraction (per attribute value) | 1,000 |
| Visual similarity encoder fine-tuning | 50,000 total |
| Outfit compatibility model | 200,000 + multi-item sets |
Get Fashion Training Data
ImageHub's fashion collection includes over 2 million on-figure, flat-lay, and cut-out images across 85 + apparel sub-categories from 30 + global retailers. Download a free fashion sample or request a targeted bundle.