Topic: Machine cognition and inference of fashion professional knowledge
Technical Area: Fashion knowledge representation and inference, Fine-grained fashion attributes classification, Mix-and-match recommendation, Fashion image synthesis, Fashion dataset
Fashion industry breeds tremendous economy value, and it is estimated to own over $3 trillion global market capitalizations by authorities. Fashion categories, like clothing, engage a large proportion of GMV for Alibaba Group, the head of global e-commerce. However, basic apparel attributes, as the key information of online clothing operation and shopping guide, rely largely on shopkeepers’ manual labeling and maintenance. It is undeniable that such artificially-created raw attributes has huge room for improvement in terms of professionalism and accuracy. To this end, two tasks, (1) creating an apparel attribute knowledge system that is both professional and machine-learnable, and (2) building a professional fashion dataset with large scale and high quality, are the foundations of constructing highly dependable fashion-related machine cognition algorithms. AI algorithms based on these two tasks could enhance the reliability of online apparel attributes, thus leading to an improvement of operation efficiency and also helping build more exquisite fashion products.
With the above-mentioned single-product attribute cognition algorithm, we aim to expand machine’s capability in fashion field, including constructing product-product relationship (e.g., mix- and-match recommendation) and product-consumer relationship (e.g., recommending clothes that fit consumers’ body figure, hairstyle and makeup).Gradually, we are able to construct apparel- related rule and logic inference ability that comes close to average consumers’ subjective fashion aesthetics. We expect that a machine can behave like a fashion expert to provide professional fashion advice for the public, like mix-and-match recommendation, as well as automatic generated comment and visualization.
Our expected collaborative teams are:
- Fashion professionals which are interested in AI applications
- AI teams with research or application experiences in fashion topics
We expect to cooperate on the following four aspects:
1. Attribute recognition
Constructing the basic recognition ability for apparel categories, like bags, shoes and accessories. In general, this process involves knowledge structure definition, dataset construction, model training and verification. Among them, knowledge structure definition is the most difficult one. There is rich connotation in fashion knowledge, so only through professional arrangement of these knowledge can it become an appropriate target for machine learning. On the premise of satisfying fashion professionalism, these knowledge should fulfill the following three requirements.
There is a clear definition for each concept, and the boundary between any two concepts in one domain is as clear as possible.
The knowledge structure of each domain covers all common products, and no concept definition is omitted.
- Visual divisibility
Product image is the carrier of abundant fashion knowledge. An appropriate granularity is often required for concept division so that different concepts can be explicitly divided via product images.
2. Matching rule
Constructing matching rule on the basis of the attributes of cloths, bags, shoes and accessories, thus enabling a machine to acquire the ability of recommending appropriate mix-and-match advice for consumers. In detail, these abilities include:
- Category matching rule
- Color matching rule
- Attribute matching rule
- Recognition ability of applicable scene, weather and seasons for given mix-and-match advice
Constructing a link between generated mix-and-match advices and consumer features on the basis of massive data and professional fashion knowledge, including:
- Estimation of expressing feeling for given mix-and-match look
- Customized mix-and-match recommendation for certain face shape, body figure makeup and hairstyle.
If a machine-generated mix-and-match can be displayed in a virtual try-on manner, it can pose great charm to consumers and facilitates their imagination of fashion highlights conveyed by the look. However, certain single-product images may not be put together on a single body due to their constrained layouts. Such limitation is expected to be avoided by artificial fashion image synthesis. Automatic visualization can even be achieved by simply appointing some abstract fashion attributes, thus circumventing the limitation of real products. The above-mentioned fashion visualization is generally expressed by a single 2D image.
Related Research Topics
List few papers related with our topics:
- DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations
- Be Your Own Prada: Fashion Synthesis with Structural Coherence
- Style2Vec: Representation Learning for Fashion Items from Style Sets
- Creating Capsule Wardrobes from Fashion Images