Frames from talks

One picture per moment where a talk shows SQLancer on a slide. Committed, because for some talks that slide is the whole of the evidence: the name is never spoken, so no transcript reaches it.

Which moments want one

Ask the dataset rather than deciding by eye:

python3 -m tools.impact.talks frames

It lists what is already captured and what is missing, and says why each missing one matters. Two cases qualify, both where the words alone are weak:

• the transcriber misheard the name — the excerpt says “SQL lenser” and the slide is where it is spelled correctly;
• nothing quotable was said — which is what a slide-only mention looks like from the transcript’s side.

A talk whose transcript was read and mentions nothing is the strongest case of the second kind. Miryung Kim’s keynote is one: 439 segments, not a word, and the name plainly there on a table of fuzzers.

Capturing one

A browser job. The trap is quality, not aim — a player drops to a small rendition whenever it seeks and upgrades only a few seconds later, so a frame grabbed as the seek lands is a blur where a table should be.

1. Open the talk at the moment: ...watch?v=<id>&t=<seconds>s.
2. Pin the quality before playing: document.querySelector('#movie_player').setPlaybackQualityRange('hd1080', 'hd1080').
3. Play, seek to the second, wait a few seconds for the stream to settle, then pause. Check video.videoWidth — 1920 means the good rendition is up.
4. Hide the player’s own furniture, which would otherwise be in the picture: .ytp-chrome-bottom, .ytp-gradient-bottom, .ytp-gradient-top, .ytp-chrome-top, .ytp-large-play-button, .ytp-pause-overlay.
5. Screenshot the <video> element’s rectangle. Take a plain screenshot first and check the page has settled — the rectangle moves while the player re-lays-out, and a stale one captures the wrong crop.
6. Read the slide back off the result. If its text cannot be read, the capture has failed, whatever it looks like at a glance.

Then file it, which trims the letterboxing, shrinks it and names it:

python3 -c "from tools.impact import talks; \
  print(talks.store_frame('<screenshot>', '<video id>', <seconds>))"

Finally add the moment to the talk’s entry in tools/impact/collectors/talks.py under frames, with a note saying what is on the slide, and re-run collect --only talks.

What a frame may say

Nothing, in words. A frame never carries an excerpt: every other excerpt in this dataset is a substring of something fetched and a check can say so, while text read off a picture is a person’s reading. The note describes what is on the slide; the picture is the evidence. The validator enforces this.