Straight answer: face enrolment needs several photos per person, not one — a straight-on shot, a couple of angled shots, and at least one taken the way the person actually looks at the gate, headgear and all. A single photo is the single biggest reason one employee's attendance keeps failing while everyone else's works fine.
If you've ever had one person who "just doesn't work" with face attendance while forty others are fine, the first thing to check isn't the camera. It's how many photos that one person was enrolled with.
Why does one photo fail?
A single photograph captures one face, from one angle, in one light. The recognition system builds its idea of "this person" from whatever it was shown. Show it only a passport-style front shot taken indoors under a tube light, and it has no idea what that same face looks like turned slightly toward a gate camera at 7 AM with the sun behind them.
People aren't static. Faces change angle as someone walks, light changes through the day, and most people wear something at the gate they weren't wearing in an HR office photo — a helmet, a turban, glasses, a mask pulled down. One photo teaches the system none of that variation, so it has to guess, and guessing under real conditions is where misses happen.
What does a good enrolment set actually look like?
Enough variation to represent how the person really appears at the point where they'll be recognised. In practice that means:
- A straight-on shot. The baseline reference.
- A left-turn and right-turn shot. People rarely walk through a gate looking dead ahead.
- The gate look. However they normally show up — turban, helmet, glasses, winter cap. If it's on their face nine days out of ten, it belongs in the enrolment.
- Gate lighting, if possible. A photo taken at the actual entrance beats a studio shot, because it matches the light the camera will see every day.
Four to six photos covering this is usually enough. It isn't about volume for its own sake — twenty near-identical photos of the same angle add little over four that cover real variation.
Does more photos always mean better accuracy?
No, and this is where enrolment drives can go wrong. Piling on dozens of photos taken in identical conditions doesn't teach the system anything new; it just takes longer to enrol and gives false confidence that "we did a thorough job." What matters is variety — different angles, different light, the accessories the person actually wears — not sheer count.
How do you fix an employee who keeps failing to match?
Look at their enrolment before you touch anything else. Most attendance failures that get blamed on "the AI not working" trace back to one thin enrolment record — often set up in a rush on someone's first day, before anyone knew what they'd typically wear or which gate they'd use.
The fix is usually straightforward: add two or three new photos taken at the actual gate, in the person's normal daily appearance, rather than deleting and starting over. Most systems support adding to an existing record. Then watch their match rate over the next few days before concluding anything deeper is wrong.
Where does this still fall short?
Even a well-built enrolment set won't fix a badly placed camera, a face turned fully away from the lens, or someone who changes their look dramatically week to week — a beard grown or shaved, glasses swapped for contacts, a scarf worn only in winter. Enrolment quality solves the "the system never learned this face properly" problem. It doesn't solve a genuinely difficult capture angle or lighting condition — that's a camera placement problem, and no amount of re-enrolling fixes it.
We build face recognition attendance systems, so take the recommendation with that in mind — but the diagnostic above works regardless of which vendor's system you're running. Before assuming your software is inaccurate, check the enrolment first. It's the cheaper fix, and it's right more often than not.