Pixelate Students' Faces in Classroom and School Video

Lesson observation where two children lack media consent

Here the tool mosaics every detected face, so it covers the two no-consent children plus everyone else. If you only need to anonymise the whole class, that's exactly the behaviour; there is no per-face on/off.

Source: lesson-obs.mp4 (1080p, 8 min)
Controls: Sample rate 6 Hz · Pixel size 18 · Padding 0.35

Progress:
  Loading face detector (one-time WebGPU + model warm-up)
  Detecting faces · sample 2881/2881
  Pixelating faces · 12 faces · 480.0s / 480.0s

Note: there is no per-face toggle -- every detected face is
mosaicked. To anonymise the whole class this is ideal.
If only specific children need covering and others must stay
visible, this engine can't do selective; use a manual
region redactor instead.

Heads-down deskwork keeps dropping the mosaic

Children working at desks tilt their faces down and away, which the near-frontal model misses. Raise the sample rate and padding to catch the moments they look up.

Problem:
  Faces tilted down -> BlazeFace short-range under-detects
  Default 4 Hz -> few of the look-up moments are sampled

Fix:
  Sample rate 8 Hz -> more frames sampled, more look-up hits
  Padding 0.40 -> wider coverage when a face is found

Reality: a face that never faces the camera may not be detected
at all -> manually redact that child before sharing.

School concert, children at a distance

A stage recording where children are small and far away. Slightly heavier pixel size keeps the blocks readable at distance; expect to manually fix distant misses.

Source: concert.mp4
Controls: Sample rate 6 Hz · Pixel size 20 · Padding 0.30

Why: small/distant faces are a weak spot for the short-range
model. pixelSize 20 keeps blocks visible; sampleHz 6 catches
children moving on stage.

Mandatory: scrub the whole clip; manually redact any
no-consent child the detector missed in the back rows.

Whole class exceeds the 12-track cap

A class of 25 produces far more than 12 clusters. The engine keeps the 12 most-detected and drops the rest, so you must confirm the no-consent children are among the covered 12.

Detected clusters: 24
Track limit: 12 (most-detected kept)
Console: face-blur - 24 clusters, capped to 12 most-detected

Result: 12 lowest-hit clusters are NOT mosaicked.
Mitigation:
  - Re-frame / split so the children needing coverage are the
    most-seen, or process the room in segments of <=12 faces
  - Always verify each no-consent child is actually mosaicked

Source is a still class photo, not a recording

This engine is the video pixelator and needs a decodable video stream. A still JPEG/PNG is rejected.

Input: class-photo.jpg
Result: error -- "Source has no video stream"
         (browser path: "Could not load source video")

Fix options:
  - Wrap the still in a 1-2s clip, then pixelate
  - Use an image/region redactor for a single class photo
  - For a name on a worksheet (non-face PII) use a region redactor

Each face crop is scaled down with nearest-neighbour sampling and scaled back up, discarding the detail — there is no key or reverse pass. For footage of minors that's the point: the copy you share has no recoverable original. Keep your un-pixelated master on the secure school drive if you ever need it.

The engine has no per-face on/off. It mosaics every face the detector finds, which is perfect for anonymising a whole class. If your scenario needs some children visible and only specific ones covered, this tool can't do selective redaction — use a manual region redactor for the specific seats instead.

Choosing the bulk face pixelator redirects to /video-tools/face-pixelate, which processes one recording and tracks faces across its frames. It is not a multi-photo uploader. Process each recording individually.

If no sampled frame yields a face, you get 'No faces detected. Try a lower sample rate or a clearer source.' In a classroom this usually means children facing away/down, small distant faces, or motion blur. Raise the sample rate, use a clearer source, and manually redact children the near-frontal model can't catch.

Detections existed but clustering produced no usable track ('Faces detected but could not be grouped into redaction regions'). Sparse single-frame hits — a child who briefly looks up once — at a low sample rate cause this. Raise the sample rate so detections accumulate into a coherent track.

After padding, clamping to the frame, and rounding to even dimensions, an edge box can shrink below the 4-pixel minimum and be skipped; if all do, you get 'Detected faces fell outside frame after clamping.' Lower the padding or reframe so children aren't flush against the border, and manually check edge seats.

Only the 12 most-detected tracks are kept (ranked by hit count); the console logs 'capped to 12 most-detected'. The most-seen children are covered; a child seen briefly can be dropped. For a full class, process in segments of 12 or fewer and confirm every no-consent child is actually mosaicked.

BlazeFace short-range is tuned for near-frontal faces. A child looking down at a desk, in hard profile, or with hair across the face is detected unreliably regardless of settings. Padding helps only after a detection exists — review the recording and manually redact any no-consent child the model skips.

The detector model, WASM graph, and (on WebGPU) shaders compile on first use, shown as 'Loading face detector (one-time WebGPU + model warm-up)'. This one-time cost happens before sampling; later runs in the same session reuse the cached detector.

Whatever the input container (.mov, .webm, etc.), the result is an H.264 .mp4 — fine for staff drives, the school channel, and portfolio submissions. There's no option to keep the original container; re-mux afterward if needed.