Re: [video-codec] Test cases with i-frames (Re: I-frame spacing (or not) (Re: Test sequences, automated and otherwise))

["Timothy B. Terriberry" <[e-mail]>]

Mo Zanaty (mzanaty) wrote:
> To clarify my all-intra comment, I did indeed mean a test config of all
> intra frames (like jct-vc), but the primary purpose of those results
> would be a separate data point for general intra efficiency in all use

I agree, particularly because there are a number of coding tools which 
one would expect to help primarily on intra frames, and it makes it much 
easier to measure their performance when not cluttered by the noise of 
two orders of magnitude more inter frames (though it's still important 
to test with inter frames to make sure they don't have some deleterious 
effect on prediction).

Personally, I think it's better to have a wide variety of still images 
on which to measure intra performance, rather than coding an existing 
video test sequence with all-intra. The main reason being that 
individual frames in a video clip are highly correlated, but the space 
of images is very large. It's a better use of testing resources to use 
uncorrelated images to explore that space as widely as possible. That's 
what informed the construction of the four still-image subsets in 
draft-daede-netvc-testing Section 5.2.