[opendtv] Re: Video compression artifacts and MPEG noise reduction

  • From: Ron Economos <w6rz@xxxxxxxxxxx>
  • To: opendtv@xxxxxxxxxxxxx
  • Date: Sun, 22 Jun 2008 18:00:53 -0700

No, it's an 8x8 /transform/. Motion estimation in H.264 can
be 4x4, 8x4, 4x8, 8x8, 16x8, 8x16 and 16x16. Here's some links:




Manfredi, Albert E wrote:

Ron Economos wrote:

BTW, the 8x8 transform (and variable quantization matrix)
was put back into H.264 High Profile after it was found that
the 4x4 transform looks like shit at HD resolutions.
Basically the 4x4 transform (with it's fixed quantization
matrix) cannot always generate enough coefficients to
provide a sharp image. Today, almost all broadcast HD H.264
is using High Profile.

You mean, just for the motion estimation, I assume. The macroblocks are
still 16 X 16.

Interesting tradeoff. The idea of gaining motion estimation efficiency
with the 4 X 4 fixed block size, because smaller parts of the image
could be used for the difference signals, had to be balanced against the
loss of high frequency components when doing the integer transform of
the motion vector. Resulting from such small blocks. If that turned out
to be a bad choice, I wonder if this has anything to do with the actual
bit rates required for H.264 HDTV, vs the more optimistic predictions
made previously?

In other words, maybe this was another reason to keep the GOP smaller
than ideal, from a compression point of view?


Other related posts: