Combing detection for 4:2:0

[sh0dan]

Written by neuron2 in another thread
Sure, I can make a guess, but I'd like to know whether you are trying to determine whether a clip as a whole has interlacing, or whether a specific frame has interlacing. If you are trying to automatically determine the chroma sampling type for a convert function to be used for the entire clip, the answer is different than if you intend to adjust on a per-frame basis. I doubt that the latter is even required, but please state your intention.

It would be for trying out "automatic" chroma upsampling mode detection for 4:2:0 colorspace conversions. If a frame is detected as combed, use interlaced chroma upsampling - otherwise framebased upsampling.

[IanB]

Perhaps correlating the chroma in Sat/Hue space with candiate subsampled luma patterns over groups of 4 lines would give interesting metrics about what chroma subsampling pattern was originally used.

i.e.

Y1 = 91 92 90 91 89 90
Y2 = 85 90 92 91 90 84
Y3 = 90 91 90 92 91 89
Y4 = 86 84 85 93 83 85

HS12 = 60/94 59/93 61/95
HS34 = 45/33 55/37 52/35

If you compare the correlation of grouping Y1/Y3 with HS12 and Y2/Y4 with HS34 against grouping Y1/Y2 with HS12 and Y3/Y4 with HS34 then the appropriate pattern should have the lower standard deviation. In my contrived example it is clearly interlaced. Testing in real life may prove this theory totally bogus, but Rome wasn't built in a day.

Regards
IanB

[Guest]

My concern about using combing detection as in Decomb is that there is no universal value for the vthresh parameter.

[sh0dan]

@neuron: Badly upsampled chroma does however not have _that_ big a visual impact on slightly interlaced images. Your sample image in your journal should be marked as interlaced (unless I'm misreading the text).

Situations like "mouth problem" issues will however still arise - but for a start even a 99.5% hitratio is something to consider.