Computer Vision News - March 2017

When first developing the algorithm, they needed to use techniques tightly coupled with hardware. They decided against using laser scanning, considering it too slow and sensitive to light. They also didn’t go for a standard color camera, which captures a couple of images in order to construct a 3D scan. Although this can yield a higher resolution, it takes more time. In a store setting, during the interaction between clerk and customer, every second counts, calling for as short of a scan time as possible. In the end, they decided to use depth cameras, in this case, Intel RealSense cameras. Similar to Microsoft Kinect, they also have multiple other sensors. This method offers faster results, though on the downside it provides much noisier data. They compensate for it with various segmentation, fitting and machine learning algorithms. The process continues to improve as they obtain more scans, allowing to train better models. They also make such a large use of deep learning, that they claim to be one of the largest deep learning deployments in retail around the world, in terms of number of stores. Alper Aydemir shares with us that during the development of Volumental, when a number of people tested out the 3D scanner, some of them were found doing unexpected things. From time to time, customers in stores would try to scan their dog’s feet! When asked why they did that, Aydemir laughs and says, “ We don’t know! That’s a good question! We would like to know that too… ” In these cases, the machine wouldn’t always work, but when it did, it would suggest very, very, very small shoes. Volumental is still a very young company at this point, but they can afford to be optimistic and excited about their product, as they continue to roll it out to many stores and use the newly gathered data to keep perfecting it. Computer Vision News Application 7 Application The largest deep learning deployments in retail around the world Volumental’s founders. Left, Executive Chairman Caroline Walerud