Frequently asked questions
Straight answers, including where the honest answer is "not yet."
How accurate is detection in poor lighting?
The pipeline applies CLAHE (Contrast-Limited Adaptive Histogram Equalization) before detection specifically to even out poor or uneven lighting, and uses a CNN-based detector rather than the faster but weaker HOG detector for better accuracy on harder conditions. We haven't published a formal accuracy benchmark across lighting conditions — what we can say concretely is which techniques are in the pipeline and why (see the About / Tech page).
Is this GDPR/BIPA compliant?
Not yet. This is a solo portfolio/demonstration project, not a commercial product or service offered to any customer — but the system does run live and continuously, matching against a small real enrolled roster, so this isn't a simulation. The face embeddings database is encrypted at rest, but no formal compliance review has been done, and encryption alone doesn't satisfy consent/notice/retention requirements on its own. See the Privacy Policy for exactly what's collected and how it's stored today.
Can it run on multiple cameras?
Yes. Each configured camera feed runs its own independent capture thread, detector, and liveness-tracking state — a blink verified on one camera can't affect another camera's matches, and every feed streams and logs alerts concurrently.
What happens to enrolled face data?
Enrolled photos are processed into face embeddings and stored locally on the server's filesystem, encrypted at rest — no cloud storage, no third-party sharing. Full detail is in the Privacy Policy.
How fast is detection?
With GPU acceleration (CUDA/cuDNN), measured at ~0.12s per frame on an RTX 4060 — about 340x faster than the same pipeline running CPU-only (~40.6s per frame), which is the difference between usable live video and not.