Outfits dataset

Task:

1、Fill in the blank

2、Outfit generation given texts or images

3、 Compatibility prediction

Dataset:

The Polyvore dataset

是一个流行的时尚搭配网站,用户可以自由搭配喜欢的时装配饰等单品,并通过在线社区分享,同时了解其它用户的穿衣搭配。Polyvore数据集是从Polyvore.com上爬取的用户搭配方案与单品信息数据集,根据爬取的数量不同有多个不同的版本。网络上公开可用的The Polyvore dataset 有 polyvore-dataset 等,其中包含的搭配Outfits方案数与单品Items数量如下:

#Outfits: 21889 (17316 for training, 1497 for validation and 3076 for testing)
#Categories: 380
#Items: 164,379
Max Items/Outfit: 8
Average Items/Outfit: 6.5
Text Available?: Titles & Descriptions

polyvore-images.tar.gz 是一个更大的版本(33,375 outfits),polyvore-dataset为其真子集。

论文中相关数据使用情况:

Learning Fashion Compatibility with Bidirectional LSTMs “只用了polyvore-dataset”
Dressing as a Whole: Outfit Compatibility Learning Based on Node-wise Graph Neural Networks “只用了polyvore-dataset”
Learning Type-Aware Embeddings for Fashion Compatibility “只用了自己爬取的更大的polyvore数据集”
Outfit Compatibility Prediction and Diagnosis with Multi-Layered Comparison Network:尚未公开论文,但从代码看就polyvore-dataset一个数据集

Learning Fashion Compatibility with Bidirectional LSTMs

将一个搭配方案视为从上到下然后到饰品的单品序列。