[opendtv] Re: Time to give up on 1080i for football
- From: Craig Birkmaier <craig@xxxxxxxxx>
- To: opendtv@xxxxxxxxxxxxx
- Date: Wed, 9 Dec 2009 07:41:34 -0500
At 3:19 PM -0500 12/8/09, John Shutt wrote:
True, the bit rates are capped so that the maximum bitrate of any
one channel is total ATSC video payload (for us about 16.7 Mbps)
minus the minimum bitrates of the other three streams. That results
in bitrate constraints of 7Mbps to 14.4Mbps for the HD service, and
800 Kbps to 7.7 Mbps for each of the three SD services.
Minimum bitrates are determined by the vintage of our Tandberg
encoders. Bits are of course further divvied up between encoders by
weighting in the stat mux.
Following up on my previous encoding post.
The major reason that VBR and stat muxing work so well is that the
information content of any video source is constantly changing. Rapid
motion can cause spikes in the information content, but there are
many other aspects of imagery that can make the bit rates spike.
There is a misguided assumption that as the resolution of the capture
device (camera) increases the information content of the captured
scenes also increases. This is only true IF the scene that the camera
is pointed at contains high frequency details that are filtered out
when captured with a lower resolution camera. And yes, thanks to
sampling theory, every camera employs filters to limit the highest
frequencies that reach the image sensor.
The major reason that SD-DVD works so well for movies is that
cinematographers go out of their way to limit the resolution when
capturing the images. Depth of field is one of the major tools used
to limit resolution, causing significant areas of the typical cinema
frame to be blurred or limited in detail. Likewise, because of the
low taking frame rate, motion blur is essential to prevent motion
discontinuity. Cinematographers are highly skilled in controlling the
motion of the camera to limit/prevent motion artifacts. Bottom line,
24P is NOT about high def detail - it is used to create a look and
feel that Hollywood strives for, which conveniently limits
resolution, minimizing the potential additional detail that can be
delivered via Blu-Ray.
There is another aspect of image capture that can make the encoding
of High Def material more difficult.
NOISE and sampling errors.
Noise is the enemy of entropy coding as it is random and cannot be
predicted - by the way, the same is true for those Hollywood types
who think it is important to capture every detail in the grain in
film.
IF the optics are held constant, as camera resolution increases, the
number of photons hitting each sensor site decreases - this directly
impacts the noise floor for the capture device. So while on one hand
the accuracy of each sample may improve as camera resolution
increases, the potential for noise and sampling errors also
increases. This is why oversampling is so important, as it allows us
to literally filter out entropy before the source gets to the encoder.
Dan noted that entropy encoding algorithms are very complex, and that
processing power impacts the complexity of the algorithms that can be
applied. The most difficult aspect of entropy coding is motion
compensated prediction. As the accuracy of the P and B frame
predictions improves, the differences that must be encoded decrease.
When you see significant coding noise and/or blocking artifacts, the
main problem is that the predictions are poor, which cause the
difference information to overwhelm the available channel capacity.
While increased processing power can improve compression algorithms,
it is important to note that high quality motion compensated
prediction is VERY DIFFICULT. It is not just a matter of tracking
motion, but also dealing with information that may not exist in ANY
frame within the encoding GOP. Here are a few of the worst
pathological cases:
1. The revealing of information that does not exist in any available
prediction frame. For example a football player who is twisting and
turning, revealing portions of his body/uniform that are not seen in
other frames.
2. Reflections and plastic deformations - bright surfaces reflect
light, creating images that must be encoded. These reflections may be
deformed by the shape of the surface. To accurately predict this
information you must know what the original scene that is being
reflected looks like as well as the geometry of the surface that is
causing the reflection. Now add motion to the reflecting surface and
you get the idea that predicting this kind of imagery takes massive
computing resources.
3. Sudden changes in lighting - any kind of strobing lights, camera
flashes, etc. can cause short transients that the encoder must deal
with.
The bottom line is that there is still a huge amount of territory to
exploit in terms of improving compression efficiency. Most encoding
algorithms still use crude block matching techniques rather than true
motion compensated prediction to create the prediction frames. The
major improvements in h.264 have more to do with improved block
matching techniques including sub pixel positioning of the blocks
that are being matched. The basic image transform is also more
efficient.
So for the moment, and probably well into the future, we can do more
to improve encoding efficiency by improving the quality of the
samples presented to the encoder than improving the encoder. This
means improved camera designs that minimize noise and reduce sampling
errors. It means oversampling to reduce the impact of entropy on the
source images. And it means that we should place more emphasis on
sample quality than the number of samples.
Bert tried to use some simple logic - i.e. A is to B is as B is to C
- to suggest that we can get more information through the channel.
Specifically that it might be possible to encode HD as efficiently as
SD.
I have a similar bit of logic relative to encoding.
In a bit rate constrained channel there is a maximum amount of
information that can be carried. If we stress the channel by trying
to carry more information than it can hold, we start to reduce the
quality of the information that reaches the decoder.
Thus it is not only possible, but highly likely, that in a bit rate
constrained channel, a high quality 480P encoding may deliver higher
quality to an HD screen than an over compressed HD version of the
same source.
This is today's reality, and the main reason why services like iTunes
are focused on sample quality rather than sample quantity.
Regards
Craig
----------------------------------------------------------------------
You can UNSUBSCRIBE from the OpenDTV list in two ways:
- Using the UNSUBSCRIBE command in your user configuration settings at FreeLists.org
- By sending a message to: opendtv-request@xxxxxxxxxxxxx with the word
unsubscribe in the subject line.
Other related posts:
