Face Recognition Guidelines
Whenever a face is detected on an IP camera or Doorkeeper kiosk, it is analysed by the Nirovision AI to assess its quality and provide a verdict on whether it matches a profile in the database.
The Nirovision AI has been designed designed and tested to meet accuracy standards within a specific context, which involves two main variables:
- The quality of the images captured via IP cameras and Doorkeeper kiosks, so faces are detected as expected.
- The quality of profiles, so matches occur as expected.
To address the first variable, we work with every customer and integration partner to ensure that cameras and kiosks are optimally placed and configured to the best possible settings. We also leverage heuristics in the Nirovision server and Doorkeeper to address scene-specific requirements and obstacles - and to improve the user experience. This is a crucial component behind the success of deployments, but not the whole story.
As for profiles, they can be created by Administrators, self-enrolled by users or synced by integrations. In all cases though, the quality of face(s) included in them directly impacts the quality of results achieved on your onsite.
1. Profiles that work
Generally, profiles do not need more than one, high quality face.
A high quality face is a sharp, bright, unoccluded (without sunglasses, masks, hats, scarves), front-facing, upright face. The best sources of high quality faces are the Doorkeeper kiosk, mobile selfies, and faces collected from a high-quality event.
Low-quality profile images include: blurry faces, zoomed-in faces, faces smaller than 100x100 pixels, distorted and stretched faces, partial and occluded faces, sideways faces.
A second, high quality quality face is needed in these scenarios:
- If a person wears glasses, it's advisable to have a profile image with glasses and one without glasses, as sometimes the glass glare interferes with recognition.
- If there are multiple kiosks and cameras connected onsite, which differ greatly in terms of lighting conditions, it's advisable to add one or two faces that illustrate significant lighting differences. The best way to diagnose this is to monitor if people struggle to be recognised on a specific camera using Camera and Similar filters of Activity.
- Sometimes there is no other option but to use a lower quality profile image (e.g. collected from a Drivers' License), and wait for better quality faces to be collected by Doorkeepers or cameras.
- Searching your activity for Events similar to this person and add a face from one of them.
- Using the Add Face Profile option.
2. Examples of high and low quality faces
✅ Sharp, front-facing, bright, unoccluded, upright faces.
⛔️ Dark images.
⛔️ Blurry faces.
⛔️ Washed out images.
⛔️ Partial faces. Avoid images with items that block the face, like sunglasses, hats or masks.
⛔️ Bad images: Extreme facial expressions.
Related Articles
Using the Import Tool
Upload up to 1000 face images at once with our Import Tool. Format your face images The maximum file size allowed is 8MB. Images must be in JPEG or PNG format. The Import Tool pre-fills every imported identity's name with its file name. We advise you ...Privacy Frequently Asked Questions
The idea of being identified by facial recognition software can make some people feel uneasy because a face is very personal to an individual. Yet facial recognition technology can be deployed securely and privately. The Data Futures Partnership ...Person Not Recognised (No Match Activity)
Nirovision uses the face(s) included in someone's profile to recognise them when they appear in front of Doorkeeper(s) and IP camera(s). That's why unrecognised activity is generally caused by low-quality profile images. Examples of low-quality ...What are profiles?
Whenever a face is detected on an IP camera or Doorkeeper kiosk, it is analysed by the Nirovision AI to assess its quality and provide a verdict on whether it matches a Profile in the database. Every Profile has a Name (or another identifier), ...Common Use Cases
Change Recognition Results Recognition is not always perfect, it is closely related to the quality of your Profiles. This document has a few pointers to help with this task: Good Quality Profiles. If you find an incorrect recognition result, the ...