SHOPACHU SAVES THE HOLIDAYS WITH CUDA-POWERED VISUAL SEARCH

by Sumit Gupta

I’ve written before about how my wife loves to shop online. Well, it’s even more true during the holidays, when she can avoid the crowds at the mall. The difficulty is, she often doesn’t know what she wants until she sees it. Now, maybe that would be fine if she had hours to burn browsing a hundred different web sites. But – especially at this time of year – who has time for that?

Well, now there’s a web site that can help shoppers who like to browse visually. Incogna’s online shopping web site, www.Shopachu.com, uses visual search to make it easier to find whatever you are looking for. Behind its simple interface are complex computer-vision algorithms that take advantage of NVIDIA’s CUDA architecture for massively parallel processing using GPUs.

To try it out, simply click a product on Shopachu’s site and select “Find Similar Products”. Shopachu shows a slew of options with similar styles. (Now my wife can continuously narrow down the search for the perfect gift for me!)

Shopachu does this remarkable feat by first aggregating products from hundreds of online retailers and then analyzing all the product images to look for certain visual properties using a concept they call “Visual Guided Navigation.” Traditional search engines such as Google and Bing index each web page using the text on it and other similar factors. But text is not very useful for indexing images. People often put incorrect descriptions and misleading labels on images (for example, “The Mean Machine” for a car), or images are not tagged at all.

Visual Guided Navigation starts by analyzing the patterns, shapes, texture and color in images to help cluster and associate similar-looking images. This process is complex and computationally intensive, with many variables that have to be taken into account, such as lighting, camera angle, background noise and photo quality. In fact, this sort of computer vision problem is still an intensively researched topic today, and Shopachu is one of the few commercial deployments to have scaled the algorithms successfully.

Shopachu’s success with computer vision is due to the massively parallel architecture of NVIDIA’s GPUs, dubbed CUDA. GPUs have hundreds of small processors that can operate concurrently on millions of pixels in an image. On a multi-core CPU an application has to look for large tasks to run on each of the CPU cores. But on a GPU, a computationally intensive task can be spread into millions of small sub-tasks (or threads) and each sub-task runs on a processor in the GPU. The result is a huge increase in speed and efficiency. Applications can then be scaled by using multiple GPUs in a server farm.

Shopachu’s Visual Guided Navigation reacts quickly to retail trends, it knows what’s popular and encourages browsing. According to the folks at Shopachu, visitors spend on average over 15 minutes browsing their vast line of products in each return session.

I encourage you try it yourself at www.shopachu.com.