An evaluation benchmark of ~18,000 garment images with structured attribute annotations for benchmarking Vision-Language Models on apparel classification.

Dataset Description

Each sample contains:

image: A garment photograph (clothing laid flat, on hangers, or worn)
prompt: Classification instruction prompt for VLM inference
ground_truth: JSON string with 9 garment attributes

Attribute Schema

Field	Description	Examples
type	Garment type	Shirt, Pants, Dress, Jacket
color	Primary color	Blue, Red, Black
pattern	Fabric pattern	Solid, Striped, Floral, Plaid
neckline	Neckline style	Crew, V-Neck, Collared Neckline
sleeve_length	Sleeve length	Short, Long, Sleeveless
closure	Closure type	Button, Zipper, Pullover
brand	Brand (if visible on tag)	Nike, Gap, N/A
size	Size (if visible on tag)	M, XL, N/A
defect_type	Visible defects	Hole, Stain, N/A

Data Sources

Images sourced from multiple collections:

crowd: Crowdsourced garment photos
posh: Resale marketplace listings
zen: Processing facility station captures
defect-spectrum: Garments with various defect types
nike-shoes-authenticity: Nike footwear samples
nike-shoes-classification: Nike footwear classification
training-eval: Cross-validation subset

Usage

Intended Use

This dataset is intended for evaluating VLM models on garment attribute extraction tasks using the Denali-AI evaluation pipeline and PeakBench. It is disjoint from the Denali-AI/train-35k training set.

Organization

Part of the Denali-AI garment classification project.

Downloads last month: 3