Aegis Vision intends to publish short technical notes on the engineering problems that matter most in school deployments. The notes are meant to be read by other practitioners and by the technical staff of institutions evaluating systems like ours. They will describe methods, data and results in enough detail to be checked, and they will state what did not work.
Planned topics
Per-site adaptation. How a detection model trained on general data is adapted to the cameras of one building, what footage is needed, who confirms it, and how to measure the change without over-fitting to the test scenes.
False-alarm reduction. How several independent checks are combined before an alarm sounds, how borderline detections are handled, and how to measure false alarms per camera per day in a way that reflects a guard's real experience.
On-premises deployment. How to run a multi-model vision system on a single server inside a school, keep it running when the internet connection drops, update it safely, and support it remotely without moving video off the site.
Evaluation practice. How to build held-out evaluation sets with staged scenes, how to report confidence intervals on small samples, and how to describe the conditions of a measurement so that it can be compared with others.
Notes
Nothing has been published yet. This list will be updated as notes are completed.
- (none)
Collaboration
We are open to working with researchers and institutions on evaluation methodology. Write to help@aegisvision.org.