[Data Contest] Summary of JD AI Fashion Challenge

Competition name: JINGdong AI Fashion Challenge

Link to the contest: fashion-challenge.github. IO

Category: Computer vision, Image classification/retrieval

Content to share: Competition introduction & thoughts of winners

The game introduction

With the expansion of China’s Fashion consumer market and the popularity of AI technology in China’s consumer sector, AI+Fashion is having a significant impact on the Fashion industry. While AI technology continues to emerge to facilitate and improve fashion-related shopping experiences, challenges remain.

Jingdong Group AI Platform and Research Institute will hold AI+Fashion Challenge competition here. The competition will launch fashion-related AI+Fashion competition, including two sub-tasks, and will hold special technical exchanges and awards in ChinaMM 2018.

Question 1: Fashion women’s style identification

Background of the problem

The judgment of clothing style is one of the most difficult multi-label classification tasks in the field of fashion combined with AI due to its strong professionalism. The competition invited professional fashion designers to mark the overall fashion style of nearly 65,000 photos of jingdong’s women’s products under their guidance, focusing on how to use AI to identify the fashion style of clothing.

The problem task

Judge one or more fashion styles of clothing according to the overall or partial design of clothing items/matches.

Given a fashion photo, the algorithms need to predict.

={Sports, leisure, OL/ commuting, Japanese, Korean, European and American, British, girl, socialite/lady, simple, natural, street/punk, ethnic}

This category is a collection of common fashion styles given by professional fashion designers according to the competition data.

The data set

Development data: including 55,000 professionally marked real display pictures of JINGdong fashion items, which are used for model training and tuning.
Test data: 10,000 real and professionally marked drawings of JINGdong fashion items.

Evaluation method

The teams will submit their results to the jingdong judging panel for f2-score based on all styles.

Winners share

To obtain buddhist scripturesteam

North branch 801team

extreme-WJLDteam

Question 2: Fashion item search

Fashion item search is an important pillar of online mall business. By analyzing the photos of fashion items taken by users, it is a technology with broad application scenarios to retrieve the corresponding products in the huge database of fashion items with high accuracy and high efficiency. It can help a number of ARTIFICIAL intelligence products such as photo shopping, personalized recommendation and advertisement click prediction. At the same time, querying the Angle, illumination, occlusion, size and quality of goods in the image brings great challenges to this problem.

Competition background

As one of the most challenging AI technologies in the field of e-commerce, photo shopping contains multiple AI algorithm modules. This competition will focus on solving one of the practical scenarios: cross-domain matching from the “user photo” of fashion items to the “e-commerce display picture”.

This competition focuses on how to solve this difference to accurately match fashion items. In order to highlight search and matching, all fashion items in this competition need to be matted according to the URL and coordinate of each item provided by the participating teams to ensure that each image contains only one fashion item in the given coordinate area.

Competition task

In the huge database of e-commerce display pictures, find the e-commerce display pictures that best match the user’s actual photos of specified fashion items.

Given a collection of e-commerce display charts of large-scale fashion items, each of them belongs to a certain category in {coat, shoes, luggage}. There is a matching set of e-commerce display pictures for any user’s actual photo of a fashion item.

In this competition, we set that there should be at least one e-commerce display map corresponding to any user’s actual photo, namely.

Given a user’s actual photo, the participating algorithm shall arrange the e-commerce display map in descending order according to the similarity, and submit the ID sequence of Top‐10 e-commerce display map. \

At the same time, we invite experts to mark and the corresponding Ground Truth e-commerce display map set scoring system will evaluate the quality of the participating algorithm according to the priority of the elements in the ID sequence.