Facial Recognition Explained
Understand how TempClock uses facial recognition to identify workers at the kiosk — and why your data stays private.
How Facial Recognition Works
The technology behind TempClock's face matching.
TempClock uses AdaVision ID, a state-of-the-art facial recognition model powered by Cloud AI powered by ONNX Runtime, hosted on UK servers. There is no special hardware required. If your device has a camera and a modern browser, facial recognition will work.
Here is what happens under the hood when the camera spots a face:
The system finds a face in the camera image and draws a bounding box around it
The face is aligned and normalised — correcting for tilt, rotation, and position so the model receives a consistent input
AdaVision ID computes a high-fidelity embedding — a rich mathematical representation of the face's unique features
The live embedding is compared against stored embeddings using cosine similarity to find a match
What is a facial embedding?
Think of a facial embedding as a unique "fingerprint" of someone's face, but expressed as a list of 128 numbers. AdaVision ID learns to map faces into a space where similar faces cluster together. The result captures the unique geometry of a person's features — eye spacing, jawline shape, nose width, and much more — far more accurate than older systems thanks to advanced training and face alignment preprocessing. No actual image is stored. It is purely mathematical.
Powered by UK-Hosted Cloud AI
No images stored — only mathematical signatures are kept.
One of the most important things to understand about TempClock's facial recognition is that all processing is handled by UK-hosted Cloud AI. The camera captures a frame, the Cloud AI processes it, and only the mathematical embedding is returned and stored.
This means:
Cloud AI processing is fast and typically completes in under a second. If your internet connection drops temporarily, the kiosk will queue clock-in entries and sync once the connection is restored.
Face Enrolment
How to register a worker's face using the guided 8-pose system.
Before a worker can use facial recognition at the kiosk, their face needs to be enrolled. TempClock uses a guided 8-pose system (similar to how a phone sets up Face ID) to capture a detailed facial profile. Each pose is quality-scored to ensure a reliable enrolment. This takes about 20 seconds and only needs to be done once per worker.
You can start face enrolment from two places:
Go to Workers, find the worker, and click the Enrol button next to their name
A flash message with an "Enrol face now" link appears after you save a new worker
Once you start enrolment, the worker stands in front of the camera and follows the on-screen guide. Here is what happens:
Front
The worker looks straight at the camera. A circular guide ring on screen shows them where to position their face. When they are in the correct position, the ring turns green. They hold still for one second and the pose is captured automatically — no button press needed. A quality score is displayed after each capture.
Turn Left
The guide ring moves to prompt the worker to turn their head to the left. They hold the position until the ring turns green and the capture happens automatically.
Turn Right
Same as the left turn, but now turning to the right. This captures the other side of the face.
Slight Left
A subtle turn to the left — less than the full turn. This captures the face at a quarter angle, which helps with real-world matching when workers approach the kiosk from the side.
Slight Right
A subtle turn to the right, matching the slight left pose. Together these quarter-angle captures significantly improve recognition from different approach directions.
Look Up
The worker tilts their head up slightly. This captures the area under the brow and chin line from a different angle.
Look Down
The worker tilts their head down. This captures the forehead and upper face from a downward angle.
Chin Down (Beard Angle)
Finally, a slightly different downward angle focused on the chin and jawline area. This is especially useful for workers with beards, as it captures the jaw and chin shape that may change with facial hair. This completes the multi-angle capture, giving the system a thorough profile of the face.
After all eight poses have been captured, the system displays an overall enrolment quality score. If quality is high enough, click Save Face Data to store the worker's facial profile. If any pose scored poorly (for example, due to blur or poor lighting), you will be prompted to re-capture that pose before saving. The worker's photo is automatically taken from the front pose and saved to their profile. From this point on, the kiosk will recognise them.
You can re-enrol a worker at any time if their appearance has changed significantly (for example, they have grown a beard or changed hairstyle). Simply repeat the enrolment process from the Workers page and the old facial data will be replaced. The 8-pose system with dedicated chin/beard angles handles facial hair much better than the previous 5-pose system.
Quality Scoring During Enrolment
Each pose is scored so you know the enrolment is reliable.
During enrolment, every pose capture is assigned a quality score from 0 to 100. The score is based on several factors: image sharpness, face size relative to the frame, lighting evenness, and whether the face is properly positioned within the guide ring.
If a pose scores below the minimum threshold, the system will ask the worker to re-capture that specific pose rather than continuing with a poor-quality image. This ensures that the stored embeddings are high quality, which directly improves day-to-day matching accuracy.
Clear image, well-lit, properly aligned. Ideal for reliable matching.
Usable but not ideal. Consider re-capturing if possible.
The system will ask you to re-capture this pose. Improve lighting or positioning.
A high-quality enrolment is the single biggest factor in reliable day-to-day matching. Spending a few extra seconds getting good quality scores during enrolment saves time every day at the kiosk.
Face Alignment Preprocessing
How AdaVision ID normalises faces before computing embeddings.
Before AdaVision ID computes an embedding, the system performs face alignment preprocessing. This step detects key facial landmarks (eyes, nose tip) and uses them to geometrically transform the face into a standardised position — correcting for head tilt, rotation, and off-centre positioning.
Alignment ensures that the same person produces consistent embeddings regardless of the exact angle they approach the kiosk. This is a significant improvement over the previous system, which was more sensitive to head angle and positioning.
Head tilt, in-plane rotation, off-centre positioning, and slight perspective differences
Aligned faces produce more consistent embeddings, which means fewer false rejections at the kiosk
Alignment happens automatically on every frame — workers do not need to do anything differently
AdaVision ID was trained on aligned faces, so this step is essential for the model to work at its best
The Matching Process
How live faces are compared to stored records.
When a worker stands in front of the kiosk, the system aligns the detected face, computes a live AdaVision-trained embedding, and then compares it against every stored embedding for workers registered at that location. The comparison uses cosine similarity, which measures the angle between two embedding vectors.
Live Face
[0.23, -0.14, 0.87, ...]
Compare
Stored Descriptors
[0.22, -0.15, 0.88, ...]
Match found
In simple terms, the system asks: "How similar are these two sets of 128 numbers?" The higher the cosine similarity, the closer the match. If the similarity exceeds a configurable threshold, the system considers it a positive match and shows the worker's name on screen. Administrators can adjust the match threshold on a per-location basis to account for challenging environments (for example, a dimly lit warehouse may benefit from a slightly lower threshold).
The result is expressed as a confidence percentage. See the Match Confidence section for details on what the different levels mean.
Multi-Frame Retry Buffer
The system tries multiple frames before giving up.
Instead of relying on a single camera frame, TempClock uses a multi-frame retry buffer. When a worker approaches the kiosk, the system captures several frames in quick succession and attempts to match each one. If the first frame does not produce a confident match (for example, due to motion blur or a brief shadow), the system automatically tries the next frame in the buffer.
This happens in the background and is invisible to the worker — they simply see the match appear on screen. The multi-frame approach significantly reduces false rejections caused by momentary environmental issues, and means workers rarely need to re-position themselves.
The system buffers up to 5 consecutive frames and tries each one until a match is found or all are exhausted
The entire multi-frame process typically completes within 1–2 seconds, so it feels instant to the worker
Each attempt is logged with its match confidence, GPS coordinates, and timestamp for audit purposes
If no frame produces a match above the threshold, the worker is prompted to try again or use PIN entry
The multi-frame retry buffer is one of the reasons AdaVision ID matching feels faster and more reliable than the previous system. Most transient issues (a passing shadow, a brief glance away) are handled automatically.
What Affects Accuracy
Factors that can help or hinder face matching.
Facial recognition is reliable in most conditions, but certain factors can affect how well the system matches a worker's face. Here are the most common ones:
Lighting
Poor lighting is the most common cause of low confidence scores. Avoid backlighting (bright window behind the worker) and very dim environments.
Angle
The worker should face the camera head-on. Looking to the side or up/down reduces the number of landmarks the system can map, lowering accuracy.
PPE & Masks
Hard hats, face masks, and balaclavas can cover key landmarks. Workers should remove face coverings briefly for recognition, or use PIN fallback.
Glasses
Regular glasses are fine. Tinted sunglasses or heavily reflective safety glasses can obscure the eye area and reduce confidence.
Distance
Workers should stand 30–60 cm from the camera. Too far away and the face is too small to map landmarks accurately. Too close and the image distorts.
Appearance Changes
Significant changes like growing or shaving a beard, dramatic weight change, or new hairstyle may require re-enrolment of the worker's face.
Tips for Best Results
Simple setup advice to get reliable recognition every time.
Follow these recommendations when setting up your kiosk and enrolling workers to get the best facial recognition performance:
The guided 8-pose enrolment with AdaVision ID captures the face from more angles than before (including dedicated beard/chin angles), which makes matching significantly more accurate. If a worker has trouble being recognised, try re-enrolling them with better lighting. Workers enrolled on the legacy system should re-enrol to take advantage of the improved AdaVision-trained embeddings.
When Matching Fails
What to do when the kiosk cannot recognise a worker.
Sometimes the kiosk will not be able to match a worker's face. This does not mean the system is broken — it usually means conditions are not ideal. Here are the most common reasons and what to do:
Too dark
The camera cannot detect enough detail. Turn on more lights or move the kiosk to a brighter area. As a temporary fix, workers can use PIN fallback.
Face covered
Masks, scarves, or heavy PPE covering the nose and mouth will prevent matching. Ask the worker to briefly lower their face covering, or use PIN entry.
Significantly different appearance
A major change in appearance (new beard, significant weight change) can affect matching. Re-enrol the worker's face from the Workers page in the manage portal.
Worker not enrolled
If a worker has never had their face enrolled, the kiosk will not recognise them. Enrol their face from the Workers page or use PIN entry for now.
When matching fails, the kiosk always offers a "Use PIN instead" link so the worker can still clock in. See the PIN Fallback guide.
Debug Kiosk for Testing
A diagnostic view that shows live match data in real time.
TempClock provides a debug kiosk that displays live match data, including the real-time confidence percentage, the number of frames being processed, and which stored embedding the system is matching against. This is invaluable for diagnosing recognition problems at a specific location.
To access the debug kiosk, append /kiosk/debug.php?location_id=X to your TempClock
URL, replacing X with the location ID. You can find the location ID in
Settings > Locations in the manage portal.
Live match percentage, embedding distance, frame count, quality scores, and the matched worker name — all updating in real time
When a worker reports persistent recognition issues, or when setting up a new kiosk location to verify conditions are good
/kiosk/debug.php?location_id=X — requires manager or admin login
The debug view is a diagnostic tool. Use the standard kiosk URL for everyday clock-ins.
The debug kiosk is especially useful when setting up a new location. Stand in front of it yourself and check that confidence scores are consistently high (85%+). If scores are low, adjust the lighting or camera position before rolling out to workers.
Self-Service Re-Enrolment
Workers can re-enrol themselves via a mobile link.
Workers no longer need to wait for a manager to re-enrol their face. TempClock now supports self-service re-enrolment via a secure mobile link. When a worker needs to update their facial data (for example, after growing a beard or getting new glasses), a manager can send them a one-time re-enrolment link by text or email.
The worker opens the link on their phone, follows the same guided 8-pose process using their phone's front camera, and their stored embeddings are updated automatically. The old data is replaced and the worker can clock in at the kiosk with their updated appearance straight away.
Go to Workers > select worker > Enrol and click Send Re-Enrolment Link. Choose text message or email.
The re-enrolment link expires after 24 hours for security. A new link can be generated at any time.
Any modern smartphone with a front camera and a browser. No app required.
The same quality scoring applies — workers will be asked to re-capture any pose that scores below the threshold.
Self-service re-enrolment is a great option for workers on remote sites or those who cannot easily visit the kiosk. The link works on any phone with a camera and a browser.
Security & Privacy
How TempClock keeps facial data safe and private.
Privacy is a top priority. Here is how TempClock protects your workers' facial data:
A note on audit snapshots
TempClock does capture a small snapshot image at the moment of clock-in for audit purposes (to verify that the right person clocked in). This image is stored securely and is separate from the facial recognition system. It is not used for matching — only for manual review if needed. You can disable audit snapshots in Settings > Privacy if your organisation does not require them.
Ready to try TempClock?
10-day free trial. All features included. No hidden fees, cancel anytime.