ld-decode Suggested improvements to ld-chroma-decoder 3D NTSC processing

Samples for improving ntsc3d. The CLD-R7G's 3dcomb does somewhat better on these examples, but are tricky for it as well.

dotcrawl on reds dotcrawl.ldf.zip ld-decode/R7G frame_ntsc_chroma_243 mpv-shot0002

artifacts on shifting colors 3dartifacts.ldf ld-decode/R7G frame_ntsc_chroma_ar43_303 mpv-shot0001

Sep 27 '20 06:09 Gamnn

The Sony logo is the same problem as the edge of the platter in dotcrawl.ldf:

frame_ntsc_chroma_118

There's a single line of colour sandwiched between two grey lines, so it's ambiguous in 2D whether it's actually colour or B/W detail - the similarity heuristic needs to be smarter about this kind of case. (If you bias it to choose the other way at the moment, then you get artefacts on colour-to-detail transitions instead.)

One possible solution would be to give the 2D decoder the ability to give an "I don't know" output when it's being used as part of the heuristic...

Sep 27 '20 11:09 atsampson

Some remaining problems in Video Essentials...

Frame 29108 - ambiguous single-line colour above "TEST":

frame_ntsc_chroma_1

Frame 30150ish - S&W test sequence - the diagonal blocks shouldn't have any chroma (the heuristic should spot that the whole signal is identical from frame to frame even though the 2D-decoded chroma isn't):

frame_ntsc_chroma_1043

Frame 36430 - for one frame after a scene transition, artefacts on hair that's almost the same colour as the previous frame:

frame_ntsc_chroma_7571

Frame 42918 - artefacts on the jacket look a bit worse in 3D than in 2D because it's otherwise sharper:

frame_ntsc_chroma_14718

Sep 27 '20 11:09 atsampson

Here's a tricky one, but not sure if anything can be done about it. They probably should have avoided this on something going to composite video. Crosshatch patterns really give the ntsc decoders trouble:

frame_ntsc_chroma_ar43_7 checker.ldf.zip