Automatic Retail Dataset Creation with Multiple Sources of Information Synchronization

We propose a novel methodology to automatically create and continuously extend a dataset of images of retail items by synchronizing two streams of information: (1) video recordings of the items being scanned at the Point of Sale (POS) and (2) the registry of the transactional database collected at the POS. Our approach is able to collect hundreds of thousands of images of distinct classes at virtually no cost, avoiding the expensive task of manual labeling at scale. Additionally, our approach is able to handle the dynamics of the retail scene such as the arrival of seasonal items and the dismiss of discontinued items, extending or reducing the image collection appropriately. Furthermore, images in our dataset are collected during the natural execution of daily operations, presenting natural illumination, occlusion, and view variations, which make it a rich dataset for robust modeling. Finally, in this paper we present the RI6K++ dataset, obtained using our methodology. The RI6K++ dataset contains over 500,000 images of 6,648 distinct labeled SKUs and can be used to develop retail computer-vision applications, such as automatic checkout, 3D reconstruction from 2D images, shelf auditing and smart cart.