The rise of online shopping has posed significant challenges for fashion retailers. The lack of ability to try on and visualize outfits has made shopping less interactive, contributing to low conversion rates and high return rates compared to brick-and-mortar shopping. Virtual dressing rooms can recreate the lost experience of trying on clothes in person. There are other companies working on a virtual dressing room. However, the reason why this is not taking off is scalability. Fashion ecommerce platforms have thousands, if not millions of SKUs. Current approaches generally require custom Photoshop work or expensive 3D models which are difficult to scale. In contrast, our solution leverages our machine learning research to automate the entire process, resulting in the first scalable virtual dressing room that can be easily integrated with any large e-commerce platform with millions of SKUs.
Rather than time-consuming 3d modeling, our system works with basic images. The goal, of course, is to produce accurate and realistic visualizations of outfits on people. A naive solution would be to simply copy-paste the garment onto the model. This presents two problems. 1) If the poses of the model/garment are mismatched, copy-paste does not work. 2) Even with ideal poses, copy-paste does not take into account garment-garment, garment-model interactions and also ignores lighting, shadows, etc. We use deep learning to overcome this problem. For problem 1) we use a series of image warpers to warp the garment onto an approximate body location in the appropriate pose. This differs from current approaches that typically use only a single warp which is extremely limited. For 2) we train an image generator that takes in relevant inputs (includes the model image, garment image, pose, etc) and produces a realistic image of the model wearing the garment. Our system produces significant improvements in size, fit, and drape compared to prior art, allowing us to create realistic images of any model wearing any combination of garments. If anyone is interested in additional details, we published an earlier version of our system here https://arxiv.org/abs/2003.10817. We also have another paper that will appear in the CVPR2021 conference soon.
This approach makes integration with retailers far easier because it requires only a single garment image on a uniform background per SKU. Upon receiving their catalog, our team processes them at a rate of 1 million images per week. We then work with the retailer to create a widget that can be easily injected into their website. The simplicity of this solution means that clients can have a virtual dressing room live in as quick as a few days. A live demo can be viewed here:https://revery.ai/demo.html.
We’ve successfully integrated with several fashion e-commerce retailers. Through working with our clients, we’ve shown that our dressing room improves the average engagement of users by 6x and, more importantly, the conversion rate by 6x. Additionally, we’ve seen increases in average order value (AOV) and decreases in return rates. Our solution also presents several use cases beyond the virtual dressing room. Because image generation is at the heart of our business, clients have also expressed interest in using our services to generate photoshoot images to forgo expensive studio photography.
Funnily enough, we have never envisioned ourselves doing a start-up in the fashion space as our backgrounds are all in computer science and research. We are all computer vision Ph.D. students at the University of Illinois at Urbana-Champaign and virtual try-on was initially just an academic pursuit. Kedan was researching fashion AI applications such as product recommendation while Min Jin was working on image generation and manipulation. Jeff was working on applied machine learning to practical problems like medicine and image search. We quickly realized that our individual expertise was compatible in tackling this difficult yet exciting problem. While image-based virtual try-on is an active research field in academia, no one has yet been able to productionize this technology. The transition from research to product is non-trivial - published research often operates on a largely simplified version of the problem. Generating realistic and accurate high-fidelity images of people and clothing is harder than it sounds. Inaccuracies are simply unacceptable for customers. People will not be happy if their miniskirt turns out to be a long skirt! It took us a year to get satisfactory results and at that point, we realized that this academic exercise can actually be a tool that real users want to use. That’s when we decided to launch Revery AI to bring a virtual dressing room shopping experience to all shoppers and retailers.
We would love to hear any feedback or answer any questions!
by Andrey
Azimov