High accuracy from open-source object detection software
YOLOv12 open-source object detection software was trained with only 700 video clips and that produced 90% accuracy for detecting guns in video footage without human review.
Summertime was school security spending season. Hopefully schools didn't get tricked into paying for expensive and outdated weapons detection systems.
Some of the biggest names in school security tech are still selling mid-2010s image classification software. The second gen of AI school security used convolutional neural networks (CNN) but the amount of compute needed to constantly inference thousands of images per second makes this financially unsustainable at scale. Even worse, companies that use a "human-in-the-loop" approach to compensate for errors have too much lag time and too much expensive human labor to be useful or sustainable (plus humans reviewing thousands of images per hour make mistakes too).
Cutting edge object detection software is now using attention-centric architecture that increases processing speed, reduces the amount of data processed, and has much higher accuracy than CNN or old school image class.
Just last week, students from University of Maryland sent me their presentation from a campus tech competition. Software that companies charge school districts millions of dollars for can be created by a couple of undergrad students. These students did some serious research because they realized that having third-party vendors accessing and storing video images from schools violates federal education privacy laws so they created a system with local storage that has an encrypted, user-permission based workflow.
Also published this week in IEEE Xplore, state-of-the-art YOLOv12 open-source object detection software was trained with only 700 video clips and that produced 90% accuracy for detecting guns in video footage without human review.
Attention architecture also helps the software just track the first person with a gun while legacy systems that individually inference every image will flag each police officer who holding a gun.
The future of AI security is bright and I hope schools haven't locked themselves into multi-year contacts with companies charging millions for products from the tech dark ages.
David Riedman is a Ph.D.c. in Artificial Intelligence studying the performance of LLM compared to human experts. Using human intelligence, he founded of the K-12 School Shooting Database, an open-source research project that documents gun violence at schools back to 1966. He hosts the weekly Back to School Shootings podcast and writes School Shooting Data Analysis & Reports on Substack.