Capturing lab screenshots without a camera

The workshop pages need pictures of the lab in use, and the claims pages need numbers. Both come from the same harness: lab.html running in a headless browser with a synthetic face in place of the webcam. The commands live in scripts-dev/README.md, under "Capture lab screenshots and measure recognition distance". This page covers why the harness exists and how to read what it produces.

The method

Chromium can be handed a video file as its webcam. The lab then runs its full pipeline against it: face-api detects, the landmark and recognition models run, the baseline is saved to local storage, the auto-find loop measures distance. Nothing is faked and nothing is drawn on afterwards. Every number in the two pictures is what the lab computed.

CODE

# 1. a still becomes a fake webcam feed (Y4M, I420, what Chromium expects)
ffmpeg -loop 1 -i face.jpeg -t 8 -r 15 \
  -vf "scale=-1:300,pad=640:480:(ow-iw)/2:(oh-ih)/2:color=0x9a938c,format=yuv420p" \
  -pix_fmt yuv420p face.y4m

# 2. serve the repo
python3 -m http.server 8127

# 3. drive the lab
node tutorials/final.js

The browser flags that matter:

CODE

--use-fake-ui-for-media-stream          auto-grant the camera prompt
--use-fake-device-for-media-stream      use a synthetic device
--use-file-for-fake-video-capture=FILE  ...fed from this Y4M
--enable-unsafe-swiftshader             WebGL without a GPU, for MediaPipe
--autoplay-policy=no-user-gesture-required

CODE

printf "file 'clean.y4m'\nfile 'painted.y4m'\n" > list.txt
ffmpeg -f concat -safe 0 -i list.txt -c copy clean-then-painted.y4m

Save the baseline while the clean segment is on screen, then read the distance once the painted segment starts. Chromium loops the file.

The script doing most of it

CODE

"capture:fixtures": "node scripts-dev/build-face-fixtures.cjs",
"capture:measure": "node scripts-dev/lab-capture.cjs measure",
"capture:shots": "node scripts-dev/lab-capture.cjs shots",
"capture:probe": "node scripts-dev/lab-capture.cjs probe",

A result worth keeping

While making these I ran all eight fixture pairs through the lab: save the baseline from the clean face, then measure the painted face in the same session, default settings, tiny_face_detector, threshold 0.58.

Pair	Painted face, peak distance	Outcome
figure1	0.23	matched
figure2	0.26	matched
figure3	0.39	matched
figure4	0.46	matched
figure5	0.57	matched, one frame short of the line
figure6	0.51	matched
figure7	0.33	matched
figure8	0.44	matched

None of the eight eludes the matcher. Every painted face was still recognised as its own baseline. figure5 reaches 0.57 against a 0.58 threshold, so it fails by one hundredth.

Conditions, because the number means nothing without them: a still image fed as a 640x480 webcam feed, even synthetic lighting, frontal pose, no motion, the vendored face-api models, default threshold. A still is the easiest possible case for a matcher, and a real face in a real room moves, so this is a floor rather than a verdict.

What it is good for: it is a documented baseline for the fixture set, and it says plainly that these eight painted looks are not yet evidence of evasion. If you want a screenshot of the "eluded" state for the workshop page, it will have to come from a look that actually crosses the line. The honest options are a real workshop recording, or painting more heavily on one of these fixtures until the distance clears 0.58 and saying so in the caption.

Why a synthetic face

Every screenshot of the lab shows a face. Using a real one means either a participant who has to consent to their face appearing in documentation that outlives the workshop, or a stock photo whose licence has to be checked each time the page is rebuilt. The pairs in tests/fixtures/synthetic-faces/ are synthetic: the same generated face twice, once bare and once with adversarial makeup applied. Nobody is depicted, so a picture can be regenerated at any time without asking anyone.

The second reason is repeatability. A person cannot hold still across eight runs. A file can, so a measurement taken today is comparable with one taken after a detector upgrade.

Important bias to know

Face recognition system works also when a person twist their face, and so a condition in real world might be harder to prove if we just use a 2D model.

Adding a figure

Drop figureN-clean.jpeg and figureN-painted.jpeg into tests/fixtures/synthetic-faces/, both showing the same face, ideally the same framing. Build the fixtures, measure, then add the row to the table above. The .y4m files are git-ignored: they are large and regenerable, so only the JPEGs and the measurements are worth committing.