Computer Vision News - June 2022

21 Liberty Defense “ These technologies have already surpassed humans as far as the ability to detect on a case-by-case basis, but we are now trying to get to levels that would be almost invisible in the image to a human. ” Dan is currently working on several innovations, including incorporating the time flow dimension into 4D labeling , and using spatial reinforcement to correlate images between different machines to combine asmuchdata as possible andmake the hard-to-see objects more apparent. “ We’re very excited about where we can go with detection and allowing, in an intelligent way, people through where they have permissible objects, ” Jeff adds, finally. “ At the end of the day, we want to make this a frictionless process. ” Part of the challenge here is creating a new workflow for efficient labeling that can overcome the problem of this firehose of data. ” The AI community is used to visible-wave (RGB) photographs and the performance yougetwithalgorithms on that typeof data. However, millimeter-wave reconstruction data behaves differently . You run the image through similar algorithms, but there are some r e c o n s t r u c t i o n artifacts because of the process. “ Algorithms trained on this millimeter-wave data are very good, ” Dan explains. “ They’re way better than people! It was a big deal when visible- wave algorithms started outperforming people, but it took a lot of work to get to that point. I’ve noticed m i l l i m e t e r - w a v e algorithms are seeing things that I have a hard time seeing in the data, which causes another problem because labeling when you can’t see something yourself, but you don’t know for certain that the signal is not there, that’s tricky! The ground truth is where the object was placed when the scan was performed. Whether or not a person can see that in the data is not as important as the fact that it is there. The algorithm will see something that people don’t see! ” Jeff agrees: