Repository navigation
_tracks: videogram_v.u1 / videogram_h.u1 are motiongrams #383
Description
_tracks._chunk_worker fills videogram_v.u1 and videogram_h.u1 from grey.mean(axis=...) where grey is the motion frame (the difference against the previous frame that also feeds qom). The result is a motiongram, not a videogram: on a concert recording it is almost black with thin vertical lines at cuts and applause, where a videogram of the same recording shows the stage lighting and the performers' positions.
Either the arrays should be named motiongram_* (with the videogram_* keys in tracks.json kept as aliases for a release), or the worker should also average the unfiltered greyscale frame into a real videogram pair. The second is more useful for timelines (render_timeline currently draws the motion version under the label "videogram_v").
Found while building a concert timeline on top of extract_tracks_parallel; worked around by taking MgVideo(proxy).videograms() on a 2 fps proxy of the video.
🤖 Generated with Claude Code
Activity