Art.com helps customers find art they love with visual search
“Life is better with art” is the motto of Art.com, a leading global online retailer selling high quality art reproductions. Their site lives up to the motto, as they possess the world's largest collection of art, with a meticulously curated catalog of images, featuring state of the art printing, molding sourced from around the world, and handcrafted custom framing services.
With such a large and diverse online catalog, product discovery is one of the central concerns in the overall online shopping experience. Art.com is working hard to provide an experience that gives customers the easiest way to discover, own, and enjoy art they love.
The Art.com category browse experience employs deep and broad taxonomy, offering views of the catalog from the perspective of art style, subject, artist, or the room to be decorated. Additionally, their keyword search system was recently overhauled to produce much more precise and relevant results. There is also a strong and continuing push to improve product filtration and sorting options for the customers by ensuring that each piece of art in the catalog has a rich description and attribution.
However, traditional category browse and keyword search discovery tools are not working as efficiently for art and images as they are in other domains, such as electronics, grocery, or apparel shopping. The issue is that art shoppers looking to decorate their homes often go after elusive notions of “style”, “feeling”, or “mood”, which are very hard to express using words for keyword search, and are also difficult to map to a category for browsing purposes. Instead, it is easier for shoppers to express their need as “something in a style” of a piece of art they already know well, or have seen elsewhere on the site.
There was therefore an obvious opportunity to expand traditional ways of product discovery with a visual search based on the style and content similarity of the art images.
Wan Agus, the SVP of Engineering, embarked on the project to create a visual search capability for Art.com. It quickly became apparent that this project would require a combination of existing subject matter and deep learning expertise. To address this need, the Art.com Search Engineering team partnered with Grid Dynamics, a global software engineering and consulting firm known for its deep expertise in online retail, product discovery, and machine learning.
The problem at hand was quite challenging and required a bit of out-the-box thinking. Imagine that you are looking to decorate your living room and you are particularly interested in one piece -- a painting in the style of old Dutch masters -- depicting a bowl of peaches:
You are looking for some more pieces of art in the same style to complete your decoration project. Now, a keyword search on “peaches” or “bowl of peaches” will return something like this:
Even though these search results are technically correct, they are not relevant for your project as they don’t capture the style of the original painting, just its subject matter. What you are really looking for are results like this:
Wouldn’t it be nice if you could just find a painting you love and then discover paintings which are visually similar to your sample? This is the exact purpose of the visual search feature.
To capture the elusive notion of painting style, Grid Dynamics engineers applied an image processing technique called “neural style transfer”. This deep learning based technique is able to blend the content of one picture with the artistic style of another, producing results which look like this:
Using this technique, the visual search team processed millions of the art images in the Art.com catalog, extracting artistic style features and searching for the most stylistically similar pieces of art.
So far, the results speak for themselves:
Kira Wampler, the CEO of Art.com, had access to the early versions of the system, and instantly recognized the power of the new feature:
“It’s very difficult to articulate preference for art in words. Most people fall in love with art when they see it. So, the work we’ve done with Grid Dynamics to develop a powerful and easy to use visual search has resulted in exactly what customers want. A way to discover art they love quickly and in a really beautiful way. And, we’re seeing it in the numbers too!” - Kira Wampler, CEO of Art.com
When rolled out to production, the visual search feature was an instant success with customers. A/B testing results showed that Art shoppers engaged with the new visual search feature on par with traditional keyword search, and those who used visual search converted up to 2.5 times better.
Never a group to rest on their laurels, the Grid Dynamics visual search team continues working hard to further improve the quality of the results and new related features, and Art.com, as ever, is looking for better ways to serve and delight art lovers all over the world.
Machine Learning and Artificial IntelligenceSearchImage similarityVisual Search