[rtcweb] Performance of rate-controlled H.264 Constrained High Profile

[Bo Burman <[e-mail]>]

Hi all,

Our recent post on comparing H.264 Constrained Baseline Profile (CBP) and VP8 objective performance with the x264 rate-control patch applied showed that H.264 CBP and VP8 are comparable.

When using that same patch applied to Constrained High Profile (CHP) settings of x264, as used in draft-burman-rtcweb-h264-proposal<https://datatracker.ietf.org/doc/draft-burman-rtcweb-h264-proposal/>, we get the following results for rate-controlled CHP:

VP8 is anchor: H.264 CHP is 16% better
H.264 is anchor: H.264 CHP is 24% better

This latter result is fully consistent with the fixed-QP tests used in the above draft.

While we still propose CBP as MTI to get maximum reach, we recommend and expect most implementations to also support CHP, which can be seen to give a clear quality advantage irrespective if rate control is used or not. Regarding technical differences between CBP and CHP, see a summary in section 8 of the above draft. There is also no difference in licensing situation between CBP and CHP.

Regarding choice of anchor; in cases like this where two competing alternatives are compared and there is no obvious anchor, it is possible to instead use a geometric mean (http://en.wikipedia.org/wiki/Geometric_mean) that will not depend on any anchor but give a single consistent figure.

Geometric mean: H.264 CHP is 22% better

/Bo