Image recognition Director
Trax Image Recognition
Yair Adato received his Ph.D. and M.Sc. in the field of Computer Vision from the Ben-Gurion University in 2012, and 2008 respectively. His thesis addresses complex motion estimation and specular shape reconstruction. He was awarded with the Google European Doctoral Fellowship 2011, and Zabey Award for best M.Sc. thesis in the Faculty of Natural Sciences, 2009. As part of his studies he attended internships at both Google and Harvard university. In the last 2 years he is working at Trax Retail Image Recognition, whereas image recognition director he is leading all the recognition related aspects.
Redefining Retail with Fine-Grained Detection
In the retail industry, auditing is a common process in which one surveys a store to collect data such as the availability of products on shelves, their assortment, visibility, pricing, promotion etc. At present, this process has largely been manual, time-consuming and very costly with results that are inaccurate (due to human error), limited (with basic statistics and KPIs) and untimely (taking weeks or months to deliver results). Furthermore, auditing a supermarket with 15,000-30,000 different classes (products), accuracy and details are complex from a recognition perspective.
Our mission in Trax Retail Image Recognition is to automate this process and as such, we have developed a fine-grained detection engine as our core technology. Unlike other visual search technologies, we are focused at the most granular level of recognition. For example, we often have to distinguish between a regular 330ml can of Coca-Cola to a limited edition, World Cup variety.
In this talk we will discuss the fine grained detection challenges and some of the solutions we have implemented. We will also share some of the challenges in creating a ground truth data collection for training and how this not a trivial task. We will illustrate the main differences between product detection and fine grained product detection as we understand it at Trax. Furthermore, we explain why trying to solve the problem using only detection algorithms in not enough, higher models of contextualization is essential. Finally, unlocking this information at large scale and world-wide, opens the path into retail’s big data and into its challenges