Contact Us
Posted by Sophia Cosby Wednesday, August, 23rd. 2017

Style Classification & Personalization through Attributes

We get into the nitty gritty of clothing ontology and attributes, looking at the automatic tagging process and what makes Fashwell's attribute classifier so damn great.

If you get excited talking about the difference between a short sleeve and a cap sleeve, or have ever made a scene about why Lacoste is not “street” then, congratulations, this article is perfect for you. We’re going to get into the nitty gritty of clothing attributes and why caring about them can add a whole lot of value to your eCommerce business.

Attribute Recognition is Fashwell’s thing. Here’s why:

Everyone has a preference for certain cuts, neck types, sleeve lengths, color schemes, patterns etc. These preferences are based on body height or shape, level of comfort or simply because they like the geometric precision of a V-neck over the unpredictability of a scoop neck. Whatever the reason, people like what they like. And they shop accordingly. For both offline and online catalogs, organizing products in terms of their physical attributes is nothing new. Scroll through any online shop, and you’ll see all the T-shirts grouped together, same with all the long sleeved shirts, the tank tops and so on. Go one level deeper and you will find all the V-neck T-shirts in one place. Filter even deeper, and you can get to all the navy blue V-neck T-shirts. It’s like magic!

Actually this is pretty standard. The magic happens when deep learning image recognition and computer vision companies like Fashwell come in and automate the entire attribute tagging process. Fashion retailers can finally say goodbye to the painstaking process of manually clicking and sorting things according to their attributes.

So how do we do it?

Fashion Product Attributes Recognition & Deep Learning

Attribute recognition and classification helps retailers find visually similar products, and it lets shoppers search with highly specific filters.

As with any part of machine learning technology, it all starts with collecting a ton of relevant training data. Fashwell specializes in fashion image recognition, so we train our neural networks exclusively on data from fashion images.

One of the main tasks facing any Fashwell engineer is teaching the network the difference between any number of products. So for the attribute classifier, it was trained to learn the qualities of a cap sleeve vs. a short sleeve, or a V-neck vs. a round neck. The actual training looks like this: The machine is fed an input image along with the desired attribute label, like “short sleeves”. The goal is for the output image to also be classified as “short sleeves”. The results are graded on a scale from 0 to 1, based on accuracy (see image below). Training continues until the machine adapts and becomes an expert on short sleeves and all other fashion attributes.

The machine is fed an input image along with the desired attribute label. The goal is for the output image to have the same label. Results are graded based on accuracy.

Fashwell’s attribute model is quite unique in its capabilities. Most image recognition models focus on a single subject, like cars, food or, as in Fashwell’s case, fashion. What sets our classifier apart is that, next to being an expert in the general field of fashion, it also specializes in specific fashion attributes: sleeve lengths, necklines, colors, patterns – even aesthetics.

Why Classifying by Style Creates Value for Fashion eCommerce Retailers

Ever since retailers have been closing more stores and focussing on eCommerce, personalization has become the name of the game. The modern consumer wants the same personalized experience online as it is in store. Although some luxury retailers are using personal shopping services (with real humans or fake ones) to meet these expectations, some fashion retailers offer personalization in other ways. They’re more subtle, yet still incredibly effective.

Take one of our clients: Spring, a fashion eCommerce marketplace based out of New York. Spring approached us to add an element of personalization to their app by sorting products according to the Spring style profiles. When visiting the Spring app, shoppers can choose between different outfits. In the backend, each outfit and each product of that outfit are labelled with a corresponding aesthetic profile. This ranges from “classic” to “bohemian” or from “street” to “heritage”. Thanks to this selection process, Spring can offer personalized recommendations to each of their shoppers.

Fashwell’s classifier was tasked with identifying the style profile of each product based on its aesthetic attributes. Just as with the physical attributes, our engineers trained the classifier on images with the relevant aesthetic attribute label. For instance, flowing peasant skirts usually equal a “bohemian” disposition, just as plaid shirts and workman boots are considered “heritage” wear. The end result is an unrivaled product: Fashwell’s attribute classifier caters to the personal style of any Spring customer and it gives Spring staff a great tool for managing their stock internally.

Fashwell’s attributes-based eCommerce solutions are fantastic for optimizing your product catalog. However you want to filter the items in your inventory, our classifier is up to the task. Want to learn more about how Fashwell is powering the future of retail? Get in touch: partnerships@fashwell.com

Previous article in Fashwell blog Next article in Fashwell blog