WEBVTT

1
00:00:00.000 --> 00:00:06.000
A 24 fps source is a timeline, not a bag of frames.

2
00:00:06.000 --> 00:00:12.000
Uniform timestamps make sparse sampling reproducible.

3
00:00:12.000 --> 00:00:18.000
Scene-aware samples add evidence around visual change.

4
00:00:18.000 --> 00:00:24.000
Every sample keeps its timestamp when frames become tensor batches.