SYNAPSIS
Tutorials  /  Deep dive · Quality check
Deep dive Script drafted ~6 min Companion to 03 · MaskBench

Quality check — did the masking work?

Structured sampling protocol, privacy vs utility checks, and QA paperwork your data steward will accept.

Video coming soon
Script drafted · Recording scheduled in batch 2 production (≈ 2027)

Learning goal

After watching, you can systematically assess a masked video on two axes — privacy protection and utility preservation — using a defensible sampling strategy, and produce documentation that a data steward or reviewer will accept.

Transcript

Cold open

Most masking failures aren't catastrophic. They're a single frame, halfway through a six-minute video, where the mask slipped. That's the one that ends up in a published clip. Don't be that person. Here's how to check.

Two questions, both required

Quality check is two questions, and you need to answer both.

One. Privacy. Is everything that needs to be hidden actually hidden? Every frame, every person, every transition.

Two. Utility. Is the information your analysis depends on still there? Hand position, gaze direction, body posture, whatever you're measuring. Aggressive masking can win the privacy check and lose the science.

Visual inspection — the right way to scrub

Don't just scrub. Dragging the timeline at full speed misses single-frame failures every time.

Use a sampling strategy. Take a frame every 30 seconds at a minimum. For a six-minute video, that's twelve checkpoints. Plus three classes of scene you always sample: every lighting change, every multi-person section, every burst of rapid motion. These are where automated processing fails.

For high-stakes data — clinical recordings, vulnerable populations, public archive submissions — get a second reviewer. Two pairs of eyes is cheap insurance.

The "would they recognise themselves?" test

There's a simple gut-check. Could someone who knows the participant — a parent, a teacher, a clinician — recognise them from this clip with no other context?

Even with the face perfectly covered, identity can leak through. Walking style. Distinctive clothing. Voice, if you've kept the audio. Background details. A piece of jewellery in every shot.

You're not aiming for zero identification risk. That's not achievable. You're aiming for acceptable for your data sensitivity and your sharing plan.

Utility — does your analysis still run?

The fastest way to check utility: run a slice of your actual analysis on the masked video.

Gesture coder — does the hand stroke land in the same frame in the masked version? If you have pose JSON from both, plot them. Major mismatches mean tracking failed; small offsets are fine. Conversation analysis — can you still see who's speaking when? Spatial relationships should be preserved even with full body masking.

Document it

Three pieces of paperwork — and this is the part data stewards actually care about.

1. Your sampling protocol. How did you choose which frames to inspect? Write it down once; reuse for every video in the study.

2. A quality-control log. What you checked, who checked it, what passed, what failed, what you did about the failures.

3. The processing log. The file MaskAnyone hands you with the download. That's the recipe. Without it, your masked video isn't reproducible.

Recap

Two axes — privacy and utility. Sample systematically, not by scrubbing. Ask whether someone who knows the participant could still recognise them. Run a slice of your analysis on the masked file to prove the science survived. And document the three things: protocol, log, parameters.

If your quality check found problems — and your first few always will — the troubleshooting deep-dive walks through the common failure modes and how to fix them.