VSD Fashion Dataset

Description

A dataset for visual similarity tasks, that includes various zero-shot and trainable tasks from the VSD Paper. Visual similarity measures the perceptual agreement between two objects based on their visual appearance. Two objects can be similar or dissimilar based on their color, shape, size, pattern, utility, and more. In fact, all of these factors and many others take part in determining the degree of visual similarity between two objects with varying importance. Therefore, defining the perceived visual simi- larity based on these factors is challenging. Nonetheless, learning visual similarities is a key building block for many practical utilities such as search, recommendations, etc.

Repository: https://github.com/vsd-benchmark/vsd
Paper: https://arxiv.org/abs/2308.14753
Leaderboard: https://vsd-benchmark.github.io/vsd/

Dataset Summary

Visual similarity measures the perceptual agreement between two objects based on their visual appearance. Two objects can be similar or dissimilar based on their color, shape, size, pattern, utility, and more. In fact, all of these factors and many others take part in determining the degree of visual similarity between two objects with varying importance. Therefore, defining the perceived visual simi- larity based on these factors is challenging. Nonetheless, learning visual similarities is a key building block for many practical utilities such as search, recommendations, etc.

Supported Tasks and Leaderboards

The dataset contains several tasks (displayed in the leaderboard site):

VSD Fashion Dataset a. In Catalog Retrieval
1. Zero Shot Retrieval Task. (in_catalog_retrieval_zero_shot)
2. Open Catalog Training Retrieval Task - Same queries appear in train and test. (in_catalog_open_catalog)
3. Closed Catalog Training Retrieval Task - Queries in train and test do not intersect. (in_catalog_closed_catalog) b. Consumer-Catalog (Wild) Retrieval
4. Zero Shot Retrieval Task catalog_wild_zero_shot (consumer-catalog_wild_zero_shot)

More information about each task can be found in the leaderboard and the paper.

For information on how to report your model metrics to our leaderboard check out our example model and HuggingFace's instructions.

Usage

Dataset Images

The dataset itself contains only annotations, and is based on DeepFashion dataset.

The DeepFashion dataset can be downloaded from here and extracted into a folder named 'img'.

The In-shop Clothes Retrieval Benchmark and Consumer-to-shop Clothes Retrieval Benchmark should be downloaded. There should be six folders in datasets/img after extraction:

  datasets/img/CLOTHING - Consumer-to-shop Clothes Retrieval Benchmark
  datasets/img/DRESSES - Consumer-to-shop Clothes Retrieval Benchmark
  datasets/img/TOPS - Consumer-to-shop Clothes Retrieval Benchmark
  datasets/img/TROUSERS - Consumer-to-shop Clothes Retrieval Benchmark
  datasets/img/MEN - In-shop Clothes Retrieval Benchmark
  datasets/img/WOMEN - In-shop Clothes Retrieval Benchmark

HuggingFace datasets api

In order to load the dataset annotations through HuggingFace hub, choose a task and run

  ds = load_dataset('vsd-benchmark/vsd-fashion', 'your_chosen_task', image_folder='./image_folder')

Where 'your_chosen_task' should be one of the tasks mentioned above, and './image_folder' should be the path that contains the 'img' folder.

Note that the zero-shot tasks only have a 'test' set.

Using the annotations directly

Instead of using 'datasets' loader, you can download the annotation files directly using:

    dataset_builder = load_dataset_builder('vsd-benchmark/vsd-fashion', task, image_folder=image_folder)
    splits = dataset_builder._split_generators(None)

License

VSD Fashion is licensed under MIT License.

Citation Information

@inproceedings{barkan2023efficient,
  title={Efficient Discovery and Effective Evaluation of Visual Perceptual Similarity: A Benchmark and Beyond},
  author={Barkan, Oren and Reiss, Tal and Weill, Jonathan and Katz, Ori and Hirsch, Roy and Malkiel, Itzik and Koenigstein, Noam},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={20007--20018},
  year={2023}
}