vsd-benchmark/vsd-fashion

---
license: mit
tags:
- vsd
pretty_name: VSD Fashion
---

# 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):

1. 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
1. 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](https://huggingface.co/vsd-benchmark/vsd_example/blob/main/README.md) and [HuggingFace's instructions](https://huggingface.co/docs/hub/model-cards#evaluation-results).

### Usage

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

The DeepFashion dataset can be downloaded from [here](https://mmlab.ie.cuhk.edu.hk/projects/DeepFashion.html) 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
```python
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:
```python
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

```
@misc{barkan2023efficient,
title={Efficient Discovery and Effective Evaluation of Visual Perceptual Similarity: A Benchmark and Beyond},
author={Oren Barkan and Tal Reiss and Jonathan Weill and Ori Katz and Roy Hirsch and Itzik Malkiel and Noam Koenigstein},
year={2023},
eprint={2308.14753},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```

原始信息汇总

VSD Fashion Dataset 概述

数据集描述

VSD Fashion Dataset 是一个用于视觉相似性任务的数据集，包含多种零样本和可训练任务。视觉相似性衡量两个物体基于视觉外观的感知一致性，涉及颜色、形状、大小、图案、用途等多个因素。

数据集总结

该数据集侧重于视觉相似性的测量，通过多个维度如颜色、形状等评估物体间的相似度，是构建搜索、推荐等实用工具的关键组成部分。

支持的任务和排行榜

数据集包含以下任务：

1. 目录检索（Catalog Retrieval）
   • 零样本检索任务（in_catalog_retrieval_zero_shot）
   • 开放目录训练检索任务（in_catalog_open_catalog）
   • 封闭目录训练检索任务（in_catalog_closed_catalog）
2. 消费者-目录（Wild）检索
   • 零样本检索任务（consumer-catalog_wild_zero_shot）

数据集使用

数据集图像

数据集本身仅包含注释，基于DeepFashion数据集。用户需从指定链接下载DeepFashion数据集，并按要求组织文件夹结构。

HuggingFace datasets api

用户可通过HuggingFace hub加载数据集注释，需指定任务和图像文件夹路径。

直接使用注释

用户也可直接下载注释文件，通过特定代码加载数据集。

许可证

VSD Fashion 数据集根据MIT许可证发布。

引用信息

@misc{barkan2023efficient, title={Efficient Discovery and Effective Evaluation of Visual Perceptual Similarity: A Benchmark and Beyond}, author={Oren Barkan and Tal Reiss and Jonathan Weill and Ori Katz and Roy Hirsch and Itzik Malkiel and Noam Koenigstein}, year={2023}, eprint={2308.14753}, archivePrefix={arXiv}, primaryClass={cs.CV} }