Why Video Is the Hard One

A document sits still. Video does not.

A face turns away and comes back. A plate is legible for eleven frames out of four hundred. A bystander walks behind a pillar and reappears as what the detector reasonably believes is a different person. Miss one frame and the release is a disclosure.

That is why the unit of work here is a track, not a frame. A detection is stored as timed data with its positions over time, so a correction applies to the object's whole appearance rather than being redrawn on every frame.

What You Can Redact

Class group	Detected
People	Persons, faces, heads, body parts
Vehicles	Car, bus, truck, bike, boat, airplane, train
License plates	Detected and read as text, so a plate is searchable
Weapons	Firearms and other weapons
Devices and screens	Monitors, laptops, phones, and the mobile data terminal in a patrol vehicle
Identity documents	ID cards, driving licenses, passports, name tags
Tattoos	Identifying body art that survives a face redaction
Street signs	And house numbers, where an address identifies a residence
PPE	Including negative forms, so a missing item is detectable
Custom objects	Anything you draw, tracked forward from one frame

How Video Redaction Works

Analyse — detection runs at upload or on demand. Scope is selectable by duration and by class, so part of a video processes on its own.
Review — the studio shows each object as a track across the timeline. Merge duplicates, split, adjust, delete false positives, draw what was missed. On MP4 sources tracking runs in the browser, so a drawn object is followed immediately with no server round trip.
Style — blur, pixelate, or solid fill, chosen per job. Greyscale export is available where color itself is identifying.
Produce — the redaction is burned into a new rendition. The output policy decides what happens to the original.

Selective and partial redaction

Keep one designated person visible while every other face in the scene is masked. Or redact an object for part of its time on screen rather than all of it, where it only needs obscuring in one span.

Redaction can also be filtered by predicted face attributes, so a job redacts a subset of faces rather than every face in frame.

Compared With a Video Editor

Frame-by-frame editing	Redactor
Draw a box, keyframe it, repeat for every object	Detect once, correct the track, done
Audio handled in a separate tool	Spoken PII in the same pass
No record of what was removed or why	Exemption code on every redaction, printed on the output
A mask layered over the original	Burned into a new rendition; the original governed by policy
One person's judgement, unrecorded	Custody trail with user, IP and timestamp on every action

What Video Redaction Does Not Do

Detection quality follows source quality. A clear, well-lit source produces better detection than a poor one.
Client-side tracking applies to MP4 sources. Other formats track server-side.
Face-attribute filtering follows predicted attributes, which are inferences with a confidence rather than established facts about a person. The underlying prediction is administrator-controlled per jurisdiction.
Automatic redaction applies what the detector found at the configured threshold. For a consequential release, the review step is the one that catches what the threshold missed.

How Video Redaction Is Evaluated

The sequence is detect, review and correct in the studio, then burn the redaction into a new rendition.
Analysis scope is selectable by duration and by class, so part of a video redacts on its own.
Styles are blur, pixelate and solid fill, with greyscale available on export.
On MP4, tracking runs in the browser and follows a drawn object immediately.
Partial redaction covers part of an object's time on screen rather than all of it.
License plates are read as text, not only detected as objects.

Video Redaction Software