Skip to content

This repository contains the 2nd place solution for SIGIR's 2022 Fashion Outfits Challenge hosted by Farfetch.

Notifications You must be signed in to change notification settings

dpalbrecht/SIGIR-2022

Repository files navigation

SIGIR-2022

This repository contains the 2nd place solution for SIGIR's 2022 Fashion Outfits Challenge hosted by Farfetch.

Challenge Overview

The main goal of this challenge is to develop a model that is able to generate outfits for each individual product. This is a particularly challenging task, as the patterns that make an outfit acceptable are really complex. Most of the properties that make products suitable to belong to the same outfit typically do not exist as product metadata: can a floral short dress be paired with white sandals and a yellow tote bag? The answer is: I need to see the products first.

In fashion, it is the intrinsic visual details and patterns of each product in an outfit that determines the outfit's quality. Being able to properly extract such features from different products and model which of them should be matched together is the demanding challenge we are proposing here. We will be focusing the challenge on the widely known task of Fill in the Blank, that consists of predicting what the missing item of a real outfit is, out of a list of candidate products. This task represents a good approximation of the original problem, as a model that is capable of understanding what should be the style characteristics of the missing product of an outfit can be easily adapted to the generation of full outfits.

For this challenge, the participants will have access to real outfits produced at FARFETCH by stylists and fashion experts. FARFETCH is the leading platform for online luxury fashion shopping, with the biggest catalogue of luxury items in the World with more than 3 million products and more than 10 thousand brands and high-end designers.

Data

The dataset provided for the challenge consists of a list of outfits described by the products that compose it, and the FARFETCH product images and metadata, and outfits composition. This dataset is generated by FARFETCH and provided to the participants in a ready to use manner. It is composed by three data sources:

  • Product metadata: A table composed by approximately 400,000 products with product attributes, such as family, category, brand.
  • Product images: Each product will have a single image with the item photographed with a frontal view in a white background. The images do not contain any people.
  • Fashion outfits: Table with approximately 300,000 outfits.

Final Standings

Rank Participant team FITB
1 cottagecore 0.799
2 Outfit>Overfit 0.764
3 SefaMerve Research 0.680
4 ColdWheels 0.660
5 lambino 0.631
6 RATC 0.552
7 CUFE 0.411
8 ConsciousAI 0.175
9 jaydo 0.152
10 Luxury A.I. 0.089
11 uva 0.049
12 seasee 0.015

About

This repository contains the 2nd place solution for SIGIR's 2022 Fashion Outfits Challenge hosted by Farfetch.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published