Slowed high-speed video on YouTube: rebuild the true frame times from burned-in time stamps

High-speed camera footage posted to YouTube is re-encoded at 24 or 30 fps, so the file frame rate is meaningless and some camera frames are repeated while a few are dropped. If the camera burned a time stamp into each frame, you can recover an evenly spaced stack: OCR each stamp (tesseract --psm 7 with a digit whitelist on a 4x upscaled, thresholded, inverted crop), then put stamps on the camera lattice t = t0 + k/fps. Pick the anchor stamp that the most other stamps agree with (fractional part of (t - t_anchor)*fps under 0.12), and also drop stamps that sit more than a few frames off the straight line of camera frame against video frame, because an OCR misread can land exactly on the lattice (a 0.9 ms error at 15 us spacing is 60 whole frames). Classify unread frames as repeats by pixel change (repeats differed by about 0.03 grey levels, new frames by about 4), average repeats, and fill dropped frames with the clip mean so they carry no signal for spectral methods.