Each facial recognition machine appears to be great during demo. Walk up to it, it calls out your name, everyone is impressed. Then you install it at the plant entrance. Sunlight falls on the lens, half of the shift wears masks or glasses, 80 people come in 10 minutes. This is the moment that tells you whether you've bought a demo machine or an entrance machine.
This manual will tell about how face recognition machines recognize people, what technology to invest into, how they work with masks, bad lighting and queues. It will also cover regulations, privacy, price. This manual is intended for HR managers, entrepreneurs and employees in India who need to buy one. The manual will state what the vendors don't mention: no facial recognition machine is perfect.
Quick takeaways
Face attendance machine is a camera terminal which reads an employee’s face to mark his presence along with a time stamp. Face identification machine creates a numeric code for each enrolled face and matches the scanned faces with those numbers. There is no need of any card, pin number, or fingerprint reader.
Four things happen in about a second:
Two plain terms come up again later. A false accept is when the machine lets the wrong person in. A false reject is when it turns away the right one. Tighten one and the other usually loosens.
So why pick face over other methods? Mostly proxy punching. You can lend a colleague your card, but you can't lend them your face. Buddy punching shows how that loss adds up. And if you're still weighing fingerprint, iris and palm, the biometric attendance system hub guide covers all four.
2D machines use an ordinary camera. They're the cheapest, but photos fool them more easily and poor light trips them up. Infrared machines light the face themselves, so dim or mixed light barely matters. 3D machines map the shape of the face and resist spoofing best, at the highest price.
Whatever you pick, ask about presentation attack detection, the technical name for spoof resistance. TheISO/IEC 30107 standard sets out how these attacks get tested and reported. Ask whether the vendor's liveness feature was evaluated against it, and ask for the test report. A marketing line isn't evidence.
Weighing face against fingerprint instead? Read fingerprint vs face recognition attendance.
Facial recognition technology is quite accurate for clear, front-facing images in good lighting conditions. Face recognition machines struggle with masks, excessive glare, dim lighting, or peculiar angles. Today's algorithms are much more advanced in terms of glasses, beards, and masks compared to older versions, although there is a huge variance among providers.
Independent testing helps here. NIST runs theFace Recognition Vendor Test (FRVT), now continued as FRTE, and its results show a big gap between the best and weakest algorithms. Ask which algorithm sits inside the machine and whether anyone has tested it independently.
Here's what to expect on the ground:
Most people assume a rejected punch means a broken machine. Usually it's the light, or a rushed enrolment photo taken in a dark corridor. Either way, keep a backup route like a PIN or supervisor approval with an audit trail, because some days a face just won't match.
A good facial recognition device recognizes someone in less than one second. However, it takes almost three seconds for a real hit since individuals will have approached the device, made eye contact with the camera and moved back. The line at shift change is based on head count and arrival time so do the math before purchasing.
Worked example: 120 employees, one gate, 10-minute shift change
False positives in a rate of 2%: two or three individuals will go through a second scan, adding less than 30 seconds
Having two scanners reduces the longest wait by half and doubles your coverage in case one breaks down. In case you want processing time under 3 minutes, use two machines for 120 people. The general recommendation is to install one machine for every 40 to 60 workers using the same entrance.
There are times when you don’t even need the second machine. Staggering shifts by 5 to 10 minutes may solve your queuing problem without cost.
The stand-alone machine is mounted at the entry point and functions for all people passing through, without reliance on individual cellular devices. The mobile face attendance application is perfect for outdoor teams, workers in remote locations, and small business organizations that cannot afford to invest in hardware equipment.
Go with a standalone machine when:
Go with a mobile app when:
Plenty of companies end up with both, machines at the office and an app for the field. Our mobile attendance app page explains how GPS, selfie and geofence checks work together.
Face information is personal information that must be protected. Provide clear notice, obtain consent, specify the purpose, use templates not photos, encrypt and destroy upon leaving. For operations in India, abide by the DPDP Act. If operating in EU and Illinois, comply with GDPR and BIPA respectively.
A practical checklist:
Aadhaar-based authentication is a separate track, and private employers generally can't require it for attendance, so use device-based templates. This is general information, not legal advice.
Prices range between Rs 8,000 and 15,000 for 2D scanning devices, Rs 12,000 and 25,000 for infrared, and Rs 20,000 and 35,000 and above for 3D scanning machines. However, these prices are only for the machine itself. The software and maintenance services will be additional costs.
The prices may vary with the make, city and the bulk orders. Therefore, prices given should only be used as a reference.
Your 10-point checklist
It’s not the camera that performed the best in the demo but rather the one that performs well at your entrance that should be purchased. Make sure you’re using the right kind of machine for your lighting, test it out with your employees, determine the required number of machines according to the number of shift changes and sort out all the issues of consent before anything happens. To be frank, most of the time when the machine is blamed for something it’s because of the above-mentioned mistakes.
And if you are looking for an automated attendance system, there’s always attendance.ai.
It is a camera terminal that works by identifying the presence of an employee based on his/her facial image. The system works by comparing the current picture with a numeric template stored within the system.
They usually cost between Rs 8,000 and 35,000. The simplest two dimensional ones would fall on the lower end, while the infrared and 3D versions cost more. Maintenance and software are additional, and in three years they would probably add more than the hardware itself.
It performs better than before, but accuracy is still affected by the use of masks and dark lighting conditions. The infrared version works well in dark conditions, and algorithms trained with masks perform better. Perform tests on your employees before purchasing.
Simple 2D machines could be. Machines that have liveness detection, IR or depth sensors resist images and videos much more. Get the results of liveness test from the vendor regarding ISO/IEC 30107.
The machine is ideal for entrances that do not change and for teams without smart phones. The app is good for field and remote employees. Lots of organizations operate both systems.