How Occlusion Breaks Face Tracking: Hats, Hands, Masks, and Crowds

Occlusion means something blocks the face. A hand, hat brim, microphone, another person, or a pillar can remove the visual information a tracker needs.

The important review point is often when the face reappears. A mask can look fine before the obstruction and still return late or attach to someone else afterward.

Separate partial and complete obstruction

With a partial obstruction, the detector may still find the visible face region, but its size or position can change abruptly.

With complete obstruction, there may be nothing to detect. A tracker can estimate movement for a while, but it cannot observe the hidden face through the object.

Do not judge the result only while the face is hidden. Inspect the transition back into view.

Use a crossing as a simple test

Watch two people pass in front of one another. Follow the person you intended to hide before, during, and after the overlap.

In selective mode, check both outputs: the target remains hidden, and the other person does not unexpectedly receive the target's treatment.

If both people may be hidden, blur-all can reduce the importance of identity association. It still requires detection and coverage of every visible face.

Try frequency changes for brief appearances

In Unseen's Local processing, Maximum checks for faces more frequently. It is worth testing when the clear appearance after an obstruction is very short.

It will not reveal pixels behind a solid object or guarantee correct grouping. Compare the same timestamp after the rerun instead of assuming a slower setting must be correct.

Cloud processing does not use the local preset control.

Decide whether the moment is needed

A half-second hand movement during a long explanation may be easy to cut around. A sports play or documentary interaction may need the full sequence.

If the moment matters and automatic treatment remains unreliable, open Review & fix faces. Put the box back on the face on a frame before and after the obstruction, then update the frames between them; each correction becomes a reference for the tracking. Check the mask on both sides of the obstruction and avoid covering unrelated action unnecessarily.

Keep review notes tied to the obstruction

Useful notes describe the trigger: “00:26, mask returns late after hand lowers” or “01:10, treatment switches after two people cross.”

Those notes are better than asking an editor to make the blur stronger, because they identify the temporal problem.

Finally, watch the corrected export continuously. Frame inspection finds gaps, while normal playback reveals distracting jumps or a mask following the wrong motion.

For other failure types, use the missed-frame diagnosis.