Computer Vision News - June 2022

20 Computer Vision Application way and should not need to worry about unknowns. They know a person is going in, and they are looking for a threat on that person. For this, they need to know the space of threats – what weapons they might be carrying and where they may be hidden. “ The biggest bottleneck we have with CONOPS is the permutations of a training set , ” Jeff explains. “ We must put in a representation of all the shapes and sizes of people going through and the range of threats to train the AI what normal is. ” Every customer has a different concept of a threat and how different objects should be treated. The system uses targeted or switchable algorithms to solve this, but the challenge is creating permutations of scans and a training and scoring set. With supervised learning , every scan has to be marked and tagged, and thousands and thousands of scans require labeling to create a broad enough set to cover every eventuality. “ Try finding a labeling tool that can label volumetric videos – I’ve looked around, and I couldn’t, ” Dan reveals. “ We effectively have a volume that is a video. We’re collecting data at 30 frames per second when somebody walks through our machine. That’s terabytes every day to label! If that’s done naively, just on images, you’re talking about labeling orders of magnitude slower than you’re collecting. “ Unlike the autonomous driving problem where you’re dealing with reconstructed 2D images, the result of our reconstruction is a volume. We use maximum intensity projection to get a view from that, and then we use several different viewpoints of the volume. ” The holographic volume is collapsed into a 2D frame and put through a single-shot projector , but that alone is not enough, as threats can be occluded at many angles arbitrarily on the body. 3D convolutions are one way of dealing with this, but the team chose a 2D approach with a specific number of views. In autonomous driving, there is an endless range of unpredictable events, but the possibilities are more constrained in this situation. It has a specific CONOPS, or concept of operations , which means peoplehave touse the systeminaparticular