Slep/LAION-RVS-Fashion

LAION — Referred Visual Search — Fashion

Introduced in Weakly-Supervised Conditional Embedding for Referred Visual Search

CRITEO AI Lab x ENPC

Simon Lepage, Jérémie Mary, David Picard

[Paper] [Demo] [Code] [BibTeX]

---

Composition

LAION-RVS-Fashion is composed of images from :

• LAION 2B EN
• LAION 2B MULTI TRANSLATED
• LAION 1B NOLANG TRANSLATED

These images have been grouped based on extracted product IDs. Each product in the training set is composed of at least a single image (isolated product), and a complex image (scene). We added categorical metadata and BLIP2 captions to each product. Please see the samples and refer to our paper for additional details.

Split	Products	Distractors
Train	272,457	-
Valid	400	99,541
Test	2,000	2,000,014

Total number of training images : 841,718.

Samples

Categories	Neck		Lower Body
BLIP2 Captions	a scarf with multi-coloured stripes		stella pants - dark suede
Categories	Feet		Bags
BLIP2 Captions	neon green patent leather heels with studs		the burberry small leather bag is brown and leather

Attributes

• URL, WIDTH, HEIGHT, punsafe, pwatermark, language: Original LAION fields. Please refer to their repository.
• TEXT: Text originally associated with the image.
• ENG_TEXT : Translated version for MULTI/NOLANG, copy of TEXT for EN.
• TYPE: SIMPLE (isolated products), COMPLEX (scenes), PARTIAL_COMPLEX (zommed-in scenes)
• PRODUCT_ID: Product identifier, allows to group together images depicting the same product.
• INDEX_SRC: ID of parquet file originally storing this image.
• CATEGORY: Categories of the products - Bags, Feet, Hands, Head, Lower Body, Neck, Outwear, Upper Body, Waist, Whole Body for the products, and NonClothing for some distractors.
• blip2_caption1, blip2_caption2: BLIP2-FlanT5XL-generated captions.

We also release bootstrap_IDs.pkl, the file used to generate the bootstrapped results of the paper. test_subsets is composed of product IDs, while dist_{N}_subsets are row indices.

---

Citing the dataset

To cite our work, please use the following BibTeX entry :

@article{lepage2023condvit,
  title={Weakly-Supervised Conditional Embedding for Referred Visual Search},
  author={Lepage, Simon and Mary, Jérémie and Picard, David},
  journal={arXiv:2306.02928},
  year={2023}
}