Common questions
The product
What's the difference between the CCTV module and AI Insights & People? CCTV records, stores, and plays back video; AI Insights & People identifies and understands the people in that video — profiles, timelines, reports, access zones. It's a facility feature layered on the CCTV module, running on the same appliance. See the boundary.
Do I need extra hardware? No — the same Edge Processor runs both. But AI inference consumes capacity: expect roughly half the camera feeds of a VSaaS-only appliance (e.g. 8 AI-analysed 1080p feeds on the 32 GB model versus ~16 recording-only).
Does it work without a member management system? Partially. Detection, unsure profiles, zone analytics, sentiment, and demographics all work. Member matching, check-in correlation, Self Learning, and access-zone plan rules need synced members and check-ins. You can also add people manually or via "Import Members".
What does Enhanced AI Analysis actually do? Quoting the setting: "Generates richer metadata including heatmaps, path tracking, and self learning. Disabling this will also turn off automatic profile matching. Increases storage usage." Practically: leave it on if you use this module — Self Learning disappears without it.
Identification
How accurate is member identification? On the Balanced preset, a face is only confirmed at 85%+ similarity — and a recent check-in adds supporting evidence (visit-assisted matching). Borderline scores get an extra pose check rather than a guess. Accuracy in practice is dominated by member photo quality — the "Invalid Photo" filter in the People list is the best predictor of who won't be identified.
Why is someone shown as "Unsure - 9f3c21ab"? The system sees them consistently but can't tie them to a member — usually a missing or invalid member photo, or a genuine non-member (guest, trial). Options: identify them manually, fix the photo, or let Self Learning correlate them from check-ins.
Can the system confuse two people? It's engineered against it: the 85% threshold, the pose check on borderline scores, a strict 90% duplicate-prevention re-check, and Self Learning's requirement of exactly one candidate member. When a wrong match does happen it's correctable — see Troubleshooting.
What are shadow photos? "Additional face images used by the facial recognition engine to help identify this person" — separate from the synced profile photo. Some are added manually; near-perfect matches (99%+) can be auto-added by self-learning, capped at a small rotating set, so recognition survives haircuts and new glasses.
Someone's on camera but never on the timeline. Why? Check People Tracking on that camera, then the detection-threshold filter — brief appearances are hidden by default. Also: blacklisted faces never appear, by design.
Privacy
Is this legal to run? Facial recognition is regulated and the obligations are yours as operator: signage, a privacy policy, a lawful basis (typically sign-up consent), and honouring deletion requests. See Security & privacy for the checklist.
Where is face data stored? In your facility's own isolated collections, in your chosen storage region. Recognition never crosses facilities. AI metadata ages out after the configured retention (default 24 months).
Can guests with share links see people data? No. Guest share links can expose cameras and footage, but people and face-detection data are explicitly excluded from guest tokens.
How do I delete someone's data? Unidentified profiles: "Delete this person". Members: "Delete Member" in the profile's Advanced section — and remove them from your member system too, or a future sync can recreate the profile. See Security & privacy.
Are the demographics real member data? No — estimated age, gender, ethnicity, and demeanor are AI estimates from imagery, used for aggregate analytics. They never gate access, and the UI warns they "may not match the data in your member management system."
Features
What's the blacklist for — banned members? Mostly the opposite: it's the fix for false positives — posters, mannequins, TVs — detected as people every day. It drops matching faces before any processing. It does not alert anyone; it's suppression, not surveillance. See Blacklist.
What counts as an access violation? An entry "where a member's active membership plan was not permitted by the zone's access rules (excluded, or not in the allowed plans)" — including outside the zone's access hours. It flags and reports; it doesn't lock doors. See Access zones.
What do Card Sharing and Tailgating mean on the timeline? Card Sharing: a check-in happened but the detected face isn't the card's owner. Tailgating: someone on site with no corresponding check-in. Both are timeline filter chips — see People timeline.
How is the sentiment score calculated? Detected facial expressions, weighted by your Business Context (Sentiment Scoring settings) and aggregated to 0–100: Positive ≥ 70, Neutral 40–69, Negative < 40. The Gym/Fitness context stops workout strain reading as unhappiness. Use it in aggregate — zone trends, not individuals.
Why don't today's numbers show in Reports? Report snapshots generate overnight for the previous day. The timeline is the live view; Reports is the trends view.
Can I export or query the data?
Yes — "Export Members" for the people list, and the seven insights_* MCP tools for everything else (timeline, profiles, check-ins, report snapshots), ideally through your AI Agent.
Should I change the recognition thresholds? Almost never. Balanced presets fit virtually all facilities; photo quality and camera placement dominate results. Read Face recognition tuning before moving anything, change one section at a time, and judge over days.