FashionFail Dataset

The FashionFail dataset, proposed in the paper "FashionFail: Addressing Failure Cases in Fashion Object Detection and Segmentation", (also check the project page) comprises 2,495 high-resolution images (2400x2400 pixels) of products found on e-commerce websites. The dataset is divided into training, validation, and test sets, consisting of 1,344, 150, and 1,001 images, respectively.

Note: The annotations are automatically generated by foundation models. However, a human annotator reviewed each sample to ensure the accuracy of the annotations.

Download Dataset

To address concerns regarding data regulations, we share only the URLs of the images, rather than sharing the image files directly. However, we provide a simple script to facilitate dataset construction. The script initially retrieves annotation files from HuggingFace Datasets, then proceeds to download images using the URLs provided in those annotation files.

First, install the repository with:

git clone https://github.com/rizavelioglu/fashionfail.git
cd fashionfail
pip install -e .

Then, execute the following script:

python fashionfail/data/make_dataset.py

which constructs the dataset inside "~/.cache/fashionfail/". An optional argument --save_dir can be set to construct the dataset in the preferred directory.

Annotation format

We follow the annotation format of the COCO dataset. The annotations are stored in the JSON format and are organized as follows:

{
 "info"           : info,         # dict: keys are shown below
 "licenses"       : [license],    # List[dict]: keys are shown below
 "categories"     : [category],   # List[dict]: keys are shown below
 "images"         : [image],      # List[dict]: keys are shown below
 "annotations"    : [annotation], # List[dict]: keys are shown below
}

info{
  "year"          : int,
  "version"       : str,
  "description"   : str,
  "contributor"   : str,
  "url"           : str,
  "date_created"  : datetime,
}

license{
  "id"            : int,
  "name"          : str,
  "url"           : str,
}

category{
  "id"            : int,
  "name"          : str,
  "supercategory" : str,
}

image{
  "id"            : int,
  "file_name"     : str,
  "height"        : int,
  "width"         : int,
  "license"       : int,
  "original_url"  : str,
}

annotation{
  "id"            : int,
  "image_id"      : int,
  "category_id"   : int,
  "area"          : int,
  "iscrowd"       : int,         # always 0 as instances represent a single object
  "bbox"          : list[float], # [x,y,width,height]
  "segmentation"  : str,         # compressed RLE: {"size", (height, widht), "counts": str}
}

License

TL;DR: Not available for commercial use, unless the FULL source code is shared!
This project is intended solely for academic research. No commercial benefits are derived from it. All images and brands are the property of their respective owners: © adidas 2023. Annotations are licensed under Server Side Public License (SSPL)

Citation

@inproceedings{velioglu2024fashionfail,
  author    = {Velioglu, Riza and Chan, Robin and Hammer, Barbara},
  title     = {FashionFail: Addressing Failure Cases in Fashion Object Detection and Segmentation},
  journal   = {IJCNN},
  eprint    = {2404.08582},
  year      = {2024},
}