On Mon, May 13, 2024 at 06:52:19PM +0300, radu.taraibuta@gmail.com wrote: > Previous observations: > > - Inconsistent code style with other filters. (Mostly using AVFilterLink* > link instead of AVFilterLink *link). > I hope it's fine now. > > - Unrelated changes, please split trivial unrelated changes into separate > patches. > Removed trivial changes from this patch. > > - Can't tables be generated at .init/.config_props time? No point in > storing them into binary. > Done. > > - Adding extra delay is not backward compatible change, it should be > implemented properly by adding option for users to select mode: next & prev > frame or just next or prev frame. > Added legacy option to the mode parameter. > > - Could split frame clone change into earlier separate patch. > Cannot be done. It's either frame clone or 1 frame delay. > > - Where are results of improvements with accuracy so it can be confirmed? > Here are my test results with manual labeling of scene changes: > 2379 Full length movie > > Method Threshold TP FP FN Precision > Recall F > Cubic 7 2357 423 22 0.847841727 0.990752417 > 0.913742973 > Cubic 10 2297 200 82 0.919903885 0.965531736 > 0.94216571 > Cubic 12 2217 146 162 0.938214135 0.931904161 > 0.935048503 > Cubic 15 2049 101 330 0.953023256 0.861286255 > 0.904835505 > Linear 2.8 2357 1060 22 0.689786362 0.990752417 > 0.813319531 > Linear 8 2099 236 280 0.898929336 0.882303489 > 0.890538821 > Linear 10 1886 173 493 0.91597863 0.792770071 > 0.849932402 > Legacy 5 2235 1260 144 0.639484979 0.939470366 > 0.760980592 > Legacy 8 1998 414 381 0.828358209 0.839848676 > 0.83406387 > Legacy 10 1743 193 636 0.900309917 0.732660782 > 0.80787949 > > 15 HDR10Plus_PB_EAC3JOC > https://mega.nz/file/nehDka6Z#C5_OPbSZkONdOp1jRmc09C9-viDc3zMj8ZHruHcWKyA > > Method Threshold TP FP FN Precision > Recall F > Cubic 10 15 0 0 1 1 1 > Linear 5 13 1 2 0.928571429 0.866666667 > 0.896551724 > Legacy 5 12 2 3 0.857142857 0.8 > 0.827586207 > > 21 (HDR HEVC 10-bit BT.2020 24fps) Exodus Sample > https://mega.nz/file/Sfw1hDpK#ErxCOpQDVjcI1gq6ZbX3vIfdtXZompkFe0jq47EhR2o > > Method Threshold TP FP FN Precision > Recall F > Cubic 10 21 0 0 1 1 1 > Linear 4 20 0 1 1 0.952380952 > 0.975609756 > Legacy 4 19 0 2 1 0.904761905 0.95 > > 94 Bieber Grammys > https://mega.nz/#!c9dhAaKA!MG5Yi-MJNATE2_KqcnNJZCRKtTWvdjJP1NwG8Ggdw3E > > Method Threshold TP FP FN Precision > Recall F > Cubic 15 91 23 3 0.798245614 0.968085106 > 0.875 > Cubic 18 85 9 9 0.904255319 0.904255319 > 0.904255319 > Linear 7 79 49 15 0.6171875 0.840425532 > 0.711711712 > Linear 8 74 28 20 0.725490196 0.787234043 > 0.755102041 > Legacy 7 74 40 20 0.649122807 0.787234043 > 0.711538462 > Legacy 8 71 26 23 0.731958763 0.755319149 > 0.743455497 > > > Improve scene detection accuracy by comparing frame with both previous and > next frame (creates one frame delay). > Add new mode parameter and new method to compute the frame difference using > cubic square to increase the weight of small changes and new mean formula. > This improves accuracy significantly. Slightly improve performance by not > using frame clone. > Add legacy mode for backward compatibility. > > Signed-off-by: raduct > --- > doc/filters.texi | 16 ++++ > libavfilter/scene_sad.c | 151 ++++++++++++++++++++++++++++++++++ > libavfilter/scene_sad.h | 6 ++ > libavfilter/vf_scdet.c | 156 +++++++++++++++++++++++++----------- > tests/fate/filter-video.mak | 3 + > 5 files changed, 284 insertions(+), 48 deletions(-) > > diff --git a/doc/filters.texi b/doc/filters.texi > index bfa8ccec8b..53814e003b 100644 > --- a/doc/filters.texi > +++ b/doc/filters.texi > @@ -21797,6 +21797,22 @@ Default value is @code{10.}. > @item sc_pass, s > Set the flag to pass scene change frames to the next filter. Default value > is @code{0} The patch is corrupted by linebreaks: Applying: area changed: scdet filter error: corrupt patch at line 16 Patch failed at 0001 area changed: scdet filter please check the linebreak settings or attach the patch or use git send-email thx [...] -- Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB Homeopathy is like voting while filling the ballot out with transparent ink. Sometimes the outcome one wanted occurs. Rarely its worse than filling out a ballot properly.