PII ZERO / EVALUATION

Recall, precision and processing cost.

A useful pilot answers three questions: what gets redacted, what needs review and what it takes to run. Set your acceptance criteria before processing at scale.

YOUR EVALUATION SCORECARD

The measurements that matter.

Agree the sample, matching rules and acceptance criteria before a pilot. Report each result with its workload and configuration.

Recommended pilot measures · targets agreed for your collection
MeasureWhat it tells you
RecallThe share of independently labeled sensitive regions detected, reported by category under stated matching rules.
PrecisionThe share of detections that match independently labeled sensitive regions under the same matching rules.
Exposure durationHow long a sensitive region remains visible, including the longest missed sequence.
Mask coverageHow much of each sensitive region is covered and how much surrounding content is obscured.
Task utilityImpact on hands, objects, task text and, where needed, downstream model performance.
Processing capacityVideo hours processed per wall-clock hour, including decoding, rendering and verification.
Hardware footprintGPU/CPU requirements, peak memory and the configuration used for the run.
Review effort & costHuman review minutes and estimated total cost per video hour.

These are the measures for a pilot, not a statement of achieved performance. Use independent labels for detection-quality metrics; report sample size and uncertainty with the result.

PROCESSED DEMONSTRATION

A face-redaction demonstration.

OriginalPII Zero output
Two researchers. One robot workspace.

Third-person lab stock footage processed with face detection, tracking and blur. This is not first-person training footage and does not evaluate text, screens, plates or audio.

DRAG TO COMPARE ↔
Single-clip demonstration · 28 September 2026
Input8 seconds · 1280 × 720 · 25 fps
Frames processed200
Face tracks2 tracks, each present in 200 frames
Frames with redaction200 / 200
Same-detector residual check0 unredacted residual detections reported

A same-detector re-scan can repeat the original detector’s blind spots. This is not independent ground truth, a recall measurement, or a guarantee of anonymization. Frame coverage measures where a mask was applied, not whether every sensitive region was found.

PILOT PROTOCOL

Measure on your actual distribution.

Label a sample

Use a held-out sample across wearers, sites and tasks. Annotate sensitive regions independently of the detector, including short exposures and difficult frames.

Measure coverage

Report region precision and recall, localization quality and missed exposure duration by PII category. Define the matching rules and report sample sizes and uncertainty alongside results.

Measure utility & cost

Measure masking of task-relevant hands, objects and text. Time the full workflow, including decoding, detection, rendering and verification, on the intended hardware. Record human review effort separately.

The published demonstrations do not provide an independently labeled first-person robotics benchmark or a general accuracy percentage. Screen detection and recapture or liveness classification are separate tasks. The demonstrations here do not establish their accuracy or represent a validated liveness benchmark.

COVERAGE DETAILS

Clear scope.
Better decisions.

What affects detection coverage?

Small identifiers, motion blur, reflections, occlusion and oblique views can affect coverage. Text detection also depends on readability, language, enabled entities and configuration. Automated redaction requires a review policy.

What does screen-text redaction cover?

The optional policy masks recognized text inside detected phones, laptops and monitors. It does not provide whole-screen masking or validated detection of every display, including a screen filling the frame.

Does a clean re-scan prove that all PII is removed?

No. The output re-scan uses the same detectors and can repeat their blind spots. Independent annotations are needed to measure precision and recall; re-scan findings are review signals.

Is spoken PII included?

The visual demonstrations have no audio track. They do not demonstrate selective spoken-PII redaction. Define audio handling separately when scoping your collection.

Do the demos establish accuracy on robotics footage?

No. They show specific outputs on stock clips. Assess first-person footage on an independently labeled, held-out sample from your capture environment, and measure downstream task utility where required.

LET’S TALK VIDEO REDACTION

Define what success looks like.

Bring your footage requirements. We’ll work through coverage, review and deployment criteria together.

15-minute introduction. No footage upload required.