metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: brand
      dtype: string
    - name: usage
      dtype: string
    - name: condition
      dtype: int64
    - name: type
      dtype: string
    - name: category
      dtype: string
    - name: price
      dtype: string
    - name: trend
      dtype: string
    - name: colors
      dtype: string
    - name: cut
      dtype: string
    - name: pattern
      dtype: string
    - name: season
      dtype: string
    - name: text
      dtype: string
    - name: pilling
      dtype: int64
    - name: damage
      dtype: string
    - name: stains
      dtype: string
    - name: holes
      dtype: string
    - name: smell
      dtype: string
    - name: material
      dtype: string
  splits:
    - name: train
      num_bytes: 6264676808.824
      num_examples: 28248
    - name: test
      num_bytes: 281469483.32
      num_examples: 3390
  download_size: 2843670317
  dataset_size: 6546146292.144
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: cc-by-4.0
language:
  - en
  - sv
pretty_name: second-ha

Clothing Dataset for Second-Hand Fashion

This dataset contains only the front image and labels from version 3 of the following dataset released on zenodo:

Three changes were made:

Front image: Only front image is uploaded here. Back and brand image are not.
Background removal: The background from the front image was removed using BiRefNet, which only supports up to 1024x1024 images - larger images were resized. The background removal is not perfect - some artifacts remain.
Rotation: The images were rotated to have a vertical orientation. Our internal experiments showed that just re-orienting the images can boost zero-shot performance.

The following contains most of the details copied from the original source at zenodo.

Code

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("fnauman/fashion-second-hand-front-only-rgb")

# Access the training split
train_dataset = dataset["train"]

# Print basic information
print(f"Dataset size: {len(train_dataset)} images") # 28248
print(f"Features: {train_dataset.features}") # 19

# Access an example
example = train_dataset[0]

image = example["image"]
# # Display the image - notebook
# from IPython.display import display
# display(example["image"])

print(f"Brand: {example['brand']}, Category: {example['category']}")
# Output: Brand: Soc (stadium), Category: Ladies

Dataset Details

Dataset Description

The dataset originates from projects focused on the sorting of used clothes within a sorting facility. The primary objective is to classify each garment into one of several categories to determine its ultimate destination: reuse, reuse outside Sweden (export), recycling, repair, remake, or thermal waste.

The dataset has 31,638 clothing items, a massive update from the 3,000 items in version 1. The dataset collection started under the Vinnova funded project "AI for resource-efficient circular fashion" in Spring, 2022 and involves collaboration among three institutions: RISE Research Institutes of Sweden AB, Wargön Innovation AB, and Myrorna AB. The dataset has received further support through the EU project, CISUTAC (cisutac.eu).

Data collected by: Wargön Innovation AB, Myrorna AB
Curation, cleaning and release by: RISE Research Institutes of Sweden AB
Funded by: Vinnova, CISUTAC - EU Horizon
License: CC-BY 4.0

Dataset Sources

Repository: Clothing Dataset for Second-Hand Fashion

Uses

Usage prediction in sorting facilities.
Detection of attributes that are specific to used or second-hand garments: condition scale (1-5), stains, holes, etc.

Dataset Structure

The dataset contains 31,638 clothing items, each with a unique item ID in a datetime format. The items are divided into three stations: station1, station2, and station3. The station1 and station2 folders contain images and annotations from Wargön Innovation AB, while the station3 folder contains data from Myrorna AB. Each clothing item has three images and a JSON file containing annotations.
Three images are provided for each clothing item:
1. Front view.
2. Back view.
3. Brand label close-up. About 4000-5000 brand images are missing because of privacy concerns: people's hands, faces, etc. Some clothing items did not have a brand label to begin with.
Image resolutions are primarily in two sizes: 1280x720 and 1920x1080. The background of the images is a table that used a measuring tape prior to January 2023, but later images have a square grid pattern with each square measuring 10x10 cm.
Each JSON file contains a list of annotations, some of which require nuanced interpretation (see labels.py for the options):
- usage: Arguably the most critical label, usage indicates the garment's intended pathway. Options include 'Reuse,' 'Repair,' 'Remake,' 'Recycle,' 'Export' (reuse outside Sweden), and 'Energy recovery' (thermal waste). About 99% of the garments fall into the 'Reuse,' 'Export,' or 'Recycle' categories.
- trend: This field refers to the general style of the garment, not a time-dependent trend as in some other datasets (e.g., Visuelle 2.0). It might be more accurately labeled as 'style.'
- material: Material annotations are mostly based on the readings from a Near Infrared (NIR) scanner and in some cases from the garment's brand label.
- Damage-related attributes include:
  - condition (1-5 scale, 5 being the best)
  - pilling (1-5 scale, 5 meaning no pilling)
  - stains, holes, smell (each with options 'None,' 'Minor,' 'Major').
  Note: 'holes' and 'smell' were introduced after November 17th, 2022, and stains previously only had 'Yes'/'No' options. For station1 and station2, we introduced additional damage location labels to assist in damage detection:
```
  "damageimage": "back",
  "damageloc": "bottom left",
  "damage": "stain ",
  "damage2image": "front",
  "damage2loc": "None",
  "damage2": "",
  "damage3image": "back",
  "damage3loc": "bottom right",
  "damage3": "stain"
```
  Taken from labels_2024_04_05_08_47_35.json file. Additionally, we annotated a few hundred images with bounding box annotations that we aim to release at a later date.
- comments: The comments field is mostly empty, but sometimes contains important information about the garment, such as a detailed text description of the damage.
Whenever possible, ISO standards have been followed to define these attributes on a 1-5 scale (e.g., pilling).
Gold dataset: 100 garments were annotated multiple times by different annotators for annotator agreement comparisons. These 100 garments are placed inside a separate folder test100.
The data has been annotated by a group of expert second-hand sorters at Wargön Innovation AB and Myrorna AB.
Some attributes, such as price, should be considered with caution. Many distinct pricing models exist in the second-hand industry:
- Price by weight
- Price by brand and demand (similar to first-hand fashion)
- Generic pricing at a fixed value (e.g., 1 Euro or 10 SEK)
Wargön Innovation AB does not set the prices in practice and their prices are suggestive only (station1 and station2). Myrorna AB (station3), in contrast, does resale and sets the prices.

Citation

Nauman, F. (2024). Clothing Dataset for Second-Hand Fashion (Version 3) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.13788681