Short answer
Redacting workplace demonstration video means blurring or masking faces, bystanders, screens, documents and identifying marks while leaving hands, tools, parts and the work surface intact. Choose a method per region, apply it only where needed, decide who may keep unredacted masters, and measure how much hand and object visibility the redaction removed. Redaction reduces exposure; it does not by itself settle consent or permission questions.
The tension in workplace footage
Demonstration video is valuable because it shows people working: their hands, their tools, the parts they handle and the stations they work at. The same frames show faces, name badges, co-workers, customers, screens and paperwork. Redaction tries to remove the second set without damaging the first, and every method trades one against the other.
The legal questions behind redaction are covered in the procurement pages. This page is about the technical choices and how to measure their cost.
What to hide and what must survive
| Hide | Keep visible |
|---|---|
| Faces of demonstrators, co-workers and bystanders | Hands, fingers and wrists |
| Name badges, names on uniforms, distinctive tattoos | Tools and the points where they touch the work |
| Screens, documents, whiteboards and customer labels | Parts, and markings the task depends on |
| Customers and their property | The work surface and changes in object state |
| Reflections in glass, mirrors and polished metal | Station layout and fixtures |
| Speech containing names, addresses or account details | Contact sounds and machine cues, if audio is kept |
Some items sit in both columns. A part label may be needed to recognize the part and also reveal a customer or a confidential process; decide case by case using confidential business information in workplace footage.
Methods and their costs
| Method | What it does | Cost to training value | Residual risk |
|---|---|---|---|
| Blur | Smooths pixels in a region | Low if confined to faces; high if it spills onto hands | Weak settings can leave faces recognizable |
| Pixelation | Averages pixels in blocks | Similar to blur, with hard edges a model can learn | Coarse settings needed to be effective |
| Solid mask | Fills the region with a flat color | Obvious artifact; models may learn "mask means person" | Low for the region; context remains |
| Synthetic face replacement | Swaps in a generated face | Scene looks natural; adds generated content | Body, voice and context still identify |
| Crop to work area | Exports only the region of interest | Loses layout and whole-body context | Low, but hands near the edge get cut |
| Audio removal or masking | Drops or bleeps speech | Loses narration and contact sounds | Voice remains a strong identifier if kept |
| Metadata scrubbing | Removes location, serials, names in file paths | None | Remaining fields can still re-identify |
No method suits every region. A typical package masks screens and documents solidly, blurs faces with a strength validated on sample frames, removes speech, and leaves hands untouched.
Detection, tracking and review
Automated pipelines detect faces, people, screens and text, track them across frames so the redaction does not flicker, and send a sample to human review. For privacy, recall matters most: one missed face in a moving clip defeats the rest. For training value, precision matters too: every false detection near the hands blurs what you are paying for. Ask for both measures.
Viewpoint changes the problem. In egocentric footage the wearer's own face rarely appears except in reflections, while co-workers and screens pass close to the hands. In exocentric footage the demonstrator's face is in nearly every frame. Head-worn research devices can also record eye-tracking camera streams and a microphone array alongside scene video (Engel et al., 2023); those streams identify the wearer directly and need their own policy, since blurring the scene video does nothing to them.
Measuring what redaction costs
Treat redaction loss as a measured quantity with thresholds you set:
- Hand overlap. Share of hand-box area covered by redaction masks, per clip.
- Active-object overlap. Share of frames in which the object being manipulated is masked.
- Detector impact. Hand and object detection rates on redacted footage compared with a comparable reference set.
- Sample review. A stratified visual check across sites, viewpoints and tasks.
These measures depend on hand and object labels, which are specified in hand-object interaction annotation.
Illustrative redaction specification
# Illustrative only — example settings a buyer adapts; not legal guidance
redaction_spec: workplace_demo_v0.2
regions:
faces: {method: blur, strength: validated on sample, track: true}
bystanders: {method: blur_full_body, applies_to: non-consenting people}
screens_docs: {method: solid_mask, includes: monitors, tablets, paper, whiteboards}
badges: {method: solid_mask}
reflections: {method: blur, review: manual}
hands_tools: {method: none, protected: true} # masks may not overlap without review
audio:
speech: removed
contact_sounds: kept where consent covers audio
streams:
eye_tracking_images: not delivered
metadata: strip device location, serials and file-path names
qa:
face_recall_sample: buyer sets sample size and minimum recall
hand_overlap_max: buyer sets per clip
version: recorded per clip as redaction_version
masters: held by supplier under retention policy; buyer access per license only
Unredacted masters
Decide early who keeps the original footage. Common arrangements: the supplier keeps masters under a written retention policy and the buyer receives redacted copies only; or the buyer also receives masters under stricter access controls where the license and applicable law allow. If redaction methods improve later, re-processing needs the masters, so agree who may re-run the pipeline and record the redaction version on every clip. Deletion obligations are covered in the procurement cluster.
Redaction and permissions
Large public datasets apply de-identification alongside consent, not instead of it; Ego4D describes both consenting participants and de-identification procedures (Grauman et al., 2021). Workplace footage needs the same pairing.
Faces are not the only identifiers. Hands, gait, voice, tattoos, uniforms and a distinctive site can identify a person, and detailed hand or body data can raise biometric questions. Under the GDPR, biometric data is defined in Article 4(14), and biometric data processed to uniquely identify a person is a special category under Article 9(1) (European Parliament and Council, 2016). Whether redaction changes the analysis depends on jurisdiction and use, so treat it as a question for counsel, starting from worker consent and identifiable people in workplace recordings and biometric privacy laws and workplace video. This is not legal advice.
Limitations
- Redaction does not make footage consented, licensable or compliant on its own.
- Residual identifiers remain: voice if audio is kept, gait, hands, tattoos, uniforms and context. Assess re-identification risk on the delivered package, not on faces alone.
- Automated detectors miss faces during fast motion, at oblique angles and behind safety glasses.
- Over-redaction removes value fastest in close-quarters tasks such as handovers, where faces and hands share the frame.
- Generated faces add synthetic content to real footage; record their use in metadata.
- Audio carries its own consent questions, covered in workplace audio in demonstrations, and redaction is one step in curating raw workplace video into training-ready clips.
Questions to ask suppliers and preparation vendors
- Which regions are redacted (faces, bystanders, screens, documents, badges, reflections), and with which method?
- What face-detection recall was measured, on what sample, and how are misses corrected?
- How much hand and active-object area was masked, per clip?
- Is audio removed, masked or kept, and how is speech containing names handled?
- Are eye-tracking images or other wearer-identifying streams delivered, and under what policy?
- Who keeps unredacted masters, for how long and under what access controls?
- Is the redaction pipeline versioned, with the version recorded on every clip?
- Can delivered clips be re-processed if a miss is found later?
Discuss your requirements
If you need workplace demonstrations with a defined redaction standard, describe the footage, the redaction you need and the permitted uses to SourceX. SourceX reviews the requirement, checks the data and the supplier's licensing permissions, and follows up about potential sources; a request does not guarantee a match. More capture and preparation topics are on the workplace task demonstration hub and the humanoid training data guide.
Explore sourcing and licensing options with SourceX
Sources
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — consent paired with de-identification in a large egocentric dataset.
- Project Aria: A New Tool for Egocentric Multi-Modal AI Research — head-worn devices that record eye-tracking and audio streams.
- Regulation (EU) 2016/679 (General Data Protection Regulation) — definition of biometric data and its special-category status.