NYCU-IR20/fashion-dataset-restructured

Fashion Dataset Restructured

This is a restructured version of the NYCU-IR20/fashion-dataset optimized for Composed Image Retrieval tasks.

Dataset Description

Dataset Summary

The Fashion Dataset Restructured contains 1,000 composed image retrieval pairs with full product attributes and images. Each sample includes:

• A query image with attributes
• A target image to retrieve
• A difference caption describing how to modify the query
• Complete product metadata (14 attributes each)

This dataset is designed for training and evaluating Composed Image Retrieval (CIR) models that take an image + text modification as input and retrieve a target image.

Dataset Structure

The dataset is split into:

• Train: 700 samples (70%)
• Validation: 150 samples (15%)
• Test: 150 samples (15%)

Each sample contains:

{
    'query_id': str,              # ID of query image
    'target_id': str,             # ID of target image to retrieve
    'query_image': PIL.Image,     # Query image
    'target_image': PIL.Image,    # Target image
    'difference_caption': str,    # Text describing modification
    'category': str,              # Fashion category (Topwear, Bottomwear, etc.)
    'query_attributes': str,      # JSON string with 14 product attributes
    'target_attributes': str,     # JSON string with 14 product attributes
}

Product Attributes (14 fields)

Each image has the following attributes stored as JSON:

• productDisplayName - Full product name
• brandName - Brand name
• gender - Target gender (Men, Women, Boys, Girls, Unisex)
• baseColour - Primary color
• colour1, colour2 - Additional colors
• usage - Usage context (Casual, Formal, Sports, etc.)
• ageGroup - Target age group
• masterCategory - Top-level category
• subCategory - Detailed category
• fashionType - Fashion type
• brandUserProfile - Brand profile
• variantName - Product variant
• productDescriptionText - Full description

Example Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("NYCU-IR20/fashion-dataset-restructured")

# Access a sample
sample = dataset['train'][0]

print(f"Category: {sample['category']}")
print(f"Modification: {sample['difference_caption']}")
print(f"Query image size: {sample['query_image'].size}")

# Parse attributes
import json
query_attrs = json.loads(sample['query_attributes'])
print(f"Query product: {query_attrs['productDisplayName']}")
print(f"Brand: {query_attrs['brandName']}")
print(f"Color: {query_attrs['baseColour']}")

Data Fields

Field	Type	Description
query_id	string	Unique identifier for query image
target_id	string	Unique identifier for target image
query_image	image	Query image (PIL Image)
target_image	image	Target image to retrieve (PIL Image)
difference_caption	string	Text describing how to modify query
category	string	Fashion category
query_attributes	string	JSON with 14 product attributes
target_attributes	string	JSON with 14 product attributes

Dataset Statistics

• Total samples: 1,000
• Total unique products: 44,448 (with metadata)
• Image resolution: Variable (typically 1080x1440)
• Categories: Topwear, Bottomwear, Shoes, Watches, and more

Source Data

This dataset is restructured from:

• Original dataset: NYCU-IR20/fashion-dataset
• Images from the Fashion Product Images Dataset
• Captions with attribute-specific modifications

Preprocessing

The restructuring process:

1. Downloaded original CSVs and images
2. Linked query-target pairs with full metadata
3. Added complete product attributes (14 fields)
4. Split into train/val/test (70/15/15)
5. Converted to HuggingFace Datasets format with Image features

Use Cases

This dataset is ideal for:

• Composed Image Retrieval (CIR): Train models to retrieve images based on query image + text modification
• Vision-Language Models: Multi-modal understanding of fashion attributes
• Image Search: Attribute-based fashion product search
• Fashion AI: Recommendation systems, style transfer

Example Task

Given:

• Query image: A blue Nike t-shirt
• Modification text: "change brand to Adidas and color to red"

Task: Retrieve the target image (red Adidas t-shirt) from the database

Citation

If you use this dataset, please cite the original fashion dataset and this restructured version:

@dataset{fashion_dataset_restructured_2025,
  title={Fashion Dataset Restructured for Composed Image Retrieval},
  author={NYCU-IR20},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/NYCU-IR20/fashion-dataset-restructured}
}

License

Please refer to the original NYCU-IR20/fashion-dataset for licensing information.

Maintenance

Dataset created: November 2025
Maintainer: NYCU-IR20
Contact: GitHub