A world leading retail company came to Alegion with a very complex video annotation use case for loss prevention. Loss of inventory due to shoplifting, employee theft, administrative errors, and fraud costs the retail industry nearly $50 billion in 2018. According to the National Retail Security Survey, 36.5% of missing inventory was attributed to shoplifting and 33.2% was due to employee theft. According to the Loss Prevention Research Council, 58% of self-checkout shoplifters categorized stealing from a self-checkout machine as easy. With this being the case, the client’s goal was to develop a computer vision model that detected shoplifting at self-checkout.
The client came to Alegion after finding that previous vendors could not meet their quality requirements. Each video being annotated varied in complexity and the annotation guidelines often had subtleties and non-intuitive directions to navigate. In addition, there were numerous classifications needed, resulting in an average of 29 judgements per frame across hundreds of thousands of frames in hours of video footage. A few examples include localizing with bounding boxes the left and right hands, hand scanners, heads, items once they’ve been picked up, the relationships between the items and the person, and the state of the hand and items if the hand is empty or holding an item.