AI to Help You Buy: How Deep Learning Provides a Better Shopping Experience
With the holiday season at an end, it’s the perfect time to snap up a bargain or sell those Christmas presents, which, while thoughtful, aren’t your style.
And while shopping online is convenient, it’s not always easy. It can be frustrating to have something in mind that you really want, searching high and low for it, only to turn up empty handed.
Styria Digital Services from Croatia, a member of our Inception program, recently worked with Willhaben, the largest marketplace website in Austria, to make selling unwanted items easier and finding the outfit of your dreams faster.
Shop Till You Drop
Thanks to Styria’s AI image recognition technology, shoppers using Willhaben’s Fashion-Cam can take a photo of the item they want and then use it to search all of the site’s listings.
Standard approaches to visual search only give semantic matches: if you search using a picture of bright pink trousers, the results generated will be trousers, but they may not be bright pink. Styria’s offering allows visual searches to be more specific and can match, for example, a short-sleeve blue shirt with stripes to a short-sleeve blue shirt with stripes. It can even recognize specific brands without the logo featuring in the image.
Styria trained its convolutional neural networks with 5 million images over a period of two weeks on four NVIDIA TITAN X GPUs. Thanks to this deep learning application, users of the Styria-powered Fashion-Cam can receive the results of their image searches within 200ms, which is comparable to text-based search results.
Styria’s work earned it the best poster award at GTC Europe 2017, where the company’s executives also had a chance to meet NVIDIA founder and CEO Jensen Huang — and run a search for his iconic leather jacket.
Going, Going, Gone
Styria is also helping sellers have a better experience through its automatic caption API. To achieve this, the team developed object recognition models that are both hierarchical and fine-grained.
The API categorizes objects according to where there is the highest level of detail as well as the highest level of confidence. So, if you wanted to sell an NVIDIA TITAN X, for example, and the API didn’t have a high level of confidence that it’s a GPU, it would be listed under the more general category of PC component.
Typically generated in less than 100ms, the results minimize the effort and clicks required of sellers to find the best category for ad postings.
Competing computer vision APIs aren’t capable of recognizing everyday objects in very fine-grained detail, according to Marko Velic, head of data science at Styria. And the implications will be felt beyond online shopping apps.
“For the level of accuracy we were looking to achieve, it was crucial that we trained our models on NVIDIA’s hardware,” says Velic. “Now our models are a great resource for anyone looking to develop AR/VR applications or robotic systems that will be operated in everyday environments.”
Styria is one of more than 2,000 members in our Inception program, which provides members with access to technology, expertise and marketing support.