Deliver High Quality High Performance HEVC via Intel® Media Server Studio

internally modulated into multiple quantisers as needed for I, P or B (including reference and
nonreference slices)) is specified in Table 0A – 0D.
Deliver High Quality, High Performance HEVC via Intel® Media Server Studio
The Media Server Studio (MSS) HEVC software encoder supports several Target Usages (TU) that range
from TU1 to TU7, and Intel® Graphics optimized version of TU4-TU7 (TU4gacc-TU7gacc) such that TU1
offers the highest quality, and TU7gacc the highest speed. In our encoding tests, the main four TU modes
– TU1 (MSS ‘Quality’ mode), TU4 (MSS ‘Balanced’ mode), TU7 (MSS ‘Speed’ mode), and TU7gacc are
tested. Further for TU1, TU4, TU7, and TU7gacc modes, B-pyramid encoding with pyramid length set to 8
is used. Likewise, different TU modes use different number of reference pictures such as 4, 3, 2, and 2
reference pictures respectively for TU1, TU4, TU7, and TU7gacc modes.
To measure quality of an HEVC codec while many techniques such as peak signal-to-noise ratio (PSNR),
structural similarity index (SSIM), other objective quality metrics, or even full subjective quality tests can
be performed, in our own competitive quality assessment tests we have found results of most such
measures to be fairly consistent with each other as long as codecs are compared in a constant quantizer
mode. This allows comparison of actual codecs independent of bit rate control (BRC) techniques.
We now briefly describe a statistically tractable technique to compare video quality produced by the
codec being tested as compared to a reference codec. Rate Distortion (RD) characteristics for both the
codecs are computed using each codec’s 4-point PSNR/Bitrate measurements followed by use of MPEG’s
new BD rate ([5]) curve fitting procedure that generates a continuous RD curve that tightly fits the
measured points. A single measurement of ‘goodness’ of the codec being tested against the reference
codec in the form of BD rate is then computed that reflects percentage difference between the codecs.
The BD rate percentage difference if positive means that the codec being tested is worse in quality, that is
it costs ‘x’ percentage more bits to generate the same PSNR quality as the reference. The BD rate
difference measurement procedure allows a straightforward way of computing and independently
verifying quality of codec with respect to a reference codec.
Quality Evaluation Results for 4:2:0 8-bit Content
Before discussing detailed BD rate differences of each Media Server Studio TU mode wrt HM 14, we first
establish the quality/bitrate tradeoff of HM 14 that will be used as reference.
First, each video sequence of each test set (such as UHD, HD1080p, HD720p, and SD/XD) is encoded using
MPEG HEVC HM14 Reference Encoder with each of 4 quantizers as specified in Tables 0A-0D; sample
encoding scripts used are shown in Appendix C. The overall PSNR (averaged over frames) for each test
sequence for each Qp for each of luma (Y), and associated chroma components (U and V) is collected
along with the generated coding bitrate. For instance, Tables 1A-1D show the luma PSNR (chroma PSNR is
also collected but is not shown to keep tables manageable in size) and total bitrate for each test
sequence for each of 4 Qps is collected and is used in curve fitting to generate a continuous RD curve for
HM14 encoding of that test sequence.
20
Next, Media Server Studio HEVC Software Encoder in each of its 4 main modes: TU1, TU4, TU7, and
TU7gacc is utilized to perform encoding of each test sequence with each of 4 quantizers, and PSNRs of Y,
U, and V, and total bitrates are collected; example encoding scripts are shown in Appendix C.