Ever come across a photo in Better Homes & Gardens or on HGTV.com and wished you could reproduce that look, but not sure where to begin or feared it would be out of your budget? Also do you really have the time to go hunt each individual item and price compare and, ugh, do math?! Why not go to your favorite online furniture website with the largest selection and use one of their awesome tools to help you? Oops! It doesn't exist.... til now.

A year ago during a brain storming session, we came up with the idea to use Computer Vision technologies to present recommendations generated from designer photos at different price points. This idea was presented as a proposal at a previous Challenge Day/Hackathon. Now that Computer Vision technologies have matured, we are able to actually implement a version of the idea. Now finding a budget-friendly solution that mimics a designer look is as simple as dragging a slider.

Video component for submission?

https://drive.google.com/file/d/1yQX-IrUpEmbmCfnORW4-TbCBlZARQ1kh/view?usp=sharing

Who is the intended audience for this project?

Customers inspired by photos of professionally staged designer rooms but don't necessarily have the budget and time to source exact matches.

What problem is this trying to solve?

How can a customer reproduce the look of a professionally designed room on a limited budget?

What is the proposed solution?

Use Computer Vision technologies to identify products in a photo and recommend visually similar alternatives for each identified product. From this pool of recommendations, present to the customer a selection of combinations at different prices to browse.

  1. Run Object Detection on an uploaded photo, or begin with a curated list of products like Shop The Look, Room Planner, or Favorites List.
  2. For each product, run Visual Search
  3. Calculate all possible ordered combinations and sort by price
  4. Present in an intuitive UI

How does this solution benefit the intended audience (and anyone else)? What is the impact?

The benefit to the audience will be saved time and money while maintaining or increasing satisfaction with Wayfair.

Potential Measurement of impact

  • Measure increase in number of items per order
  • Measure increase in Average Order Value
  • Improve Wayfair's overall NPS thru one more useful and easy-to-use tool

No estimates for these measurements

What’s next?

  • The biggest problem to deal with is scaling. Depending on class, Visual Search returns the most visually similar matches within the first 20 results. Though often less, Object Detection can identify up to 10 products. This translates to over 10 trillion combinations to sort and search!!! How do you pare this down to a more manageable number? For the demo, the number of combinations was limited to 1024---5 products, 4 recommendations each.
  • Room Style identification is currently in development and could be applied to this problem.
  • If you're happy with one or more of the items but still unsure about the others, add the ability to lock a selection.
  • Reducing the load time and improving other performance issues.
  • There was only time to do one implementation, namely Shop The Look. Next would be a stand-alone feature that involves uploading an image and using Object Detection, very similar to how Search With Photo works. Also Favorite Lists and Room Planner integrations could be done as well.

Try it out

https://git.csnzoo.com/wayfair/resources/merge_requests/26168

Built With

Share this project:

Updates