|
I’m already fully deployed with headend receivers and set-tops– in
the millions – and they only use MPEG-2 compression. What’s in it for me
to go to MPEG-4 Advanced Video Compression (AVC)?
The bandwidth savings. Each day, more and more program networks are
announcing plans for multi-channel HDTV. The digital shelf space of even
the 860 MHz cable operator is starting to feel the strain. The intent of
MPEG-4 AVC (also known as H.264) is to decrease the bit rate of an HD
stream by a half – or more – while keeping picture quality the same as
(or better than) it was with prior compression technologies, like
MPEG-2. Of course, MPEG-2 and MPEG-4 AVC will likely co-exist for a long
time. For example, a customer with a large amount of MPEG-2 compressed
video content will need to de-compress the MPEG-2 streams before
encoding the reconstructed video using MPEG-4 AVC in order to achieve
bandwidth savings. Another example is a customer with a large deployed
base of MPEG-2 set-tops, who will still need to transmit MPEG-2 streams
in its distribution network to subscribers. Therefore,
currently-deployed MPEG-4 AVC encoders will need to have the ability to
decode MPEG-2 streams, and MPEG-4 AVC receivers will also need to have
the ability to MPEG-2 encode the reconstructed video.
What are the bandwidth gains?
You can achieve a 2:1 compression advantage or bandwidth savings using
some of the commercially-available MPEG-4 AVC encoders, compared to a
well-optimized MPEG-2 encoder. But the more important advantage is that
MPEG-4 AVC is still in its infancy. Its full potential hasn’t been
measured yet, but, MPEG-4 AVC will probably yield at least a 3:1
advantage by end of 2008. The reason is, MPEG-4 AVC brings much more
functionality, and much more freedom, in the use of encoding tools,
relative to MPEG-2. So the improvement gap, from the day the first
MPEG-4 AVC encoder showed up, to the expectation of what bandwidth
savings you’d get 10 years from now, is a lot wider than it was for
MPEG-2.
What are the most important or most notable technical differences
between MPEG-2 and MPEG-4 AVC?
An encoder built for MPEG-4 AVC compression contains the same core
elements as an encoder built for MPEG-2 compression: A intra/inter
predictor and a core encoder. That hasn’t changed. The differences are
in each of the pieces themselves. Much of the difference occurs in inter
prediction.
MPEG-4 AVC brings many new functionalities and features to motion
estimation, which is used for inter-prediction. For example, you can
order pictures in almost an arbitrary way. You can use normal MPEG-2
GOPs (groups of pictures), or hierarchical GOPs, or even more complex
GOPs. In a hierarchical GOP, a B-picture can be used as a reference
picture, allowing the possibility of a GOP without a P-picture (unlike
in MPEG-2). The use of hierarchical GOPs enables you to more closely
match the prediction structure to the video content itself.
Another flexibility is that with MPEG-4 AVC, you can use seven types of
blocks to partition a 16x16 block in many ways for motion estimation and
compensation, whereas with MPEG-2, you could use only non-partitioned
16x16 blocks. In other words, you are now able to use multiple motion
vectors to represent a 16x16 block, which in turn allows you to better
adapt your motion-compensation model to the video content. For instance,
you could employ large, 16x16 blocks for content that isn’t complex, but
for scenes that contain small elements, you might want to divide the
16x16 block into smaller blocks. MPEG-4 AVC lets you zoom in and out.
A third aspect is the use of quarter-bit motion vector accuracy in
MPEG-4 AVC, as opposed to the half-bit motion vector accuracy supported
in MPEG-2. The higher-precision motion estimation, which is very
important for very detailed/slow motion, lets you capture the motion
within the video content better. For example, with half-bit precision,
you might miss a slowly-moving object, which would yield larger
prediction errors, which means you would waste bits.
This MPEG-4 AVC benefits go on and on. There is also the use of multiple
reference pictures, which allows the motion-compensation model to
address occlusion and other effects. For example, a B-picture can now
depend on two previous, or three previous, or even multiple future
pictures.
Lastly, MPEG-4 AVC adds a weighted prediction factor. In MPEG-2, the two
predicted pixel values from the pictures previous to and subsequent to a
B-picture are simply averaged. Weighted prediction lets you add weighing
factors as well as offsets to the predicted pixel values. This delivers
very impressive performance by an MPEG-4 AVC as it handles video special
effects such as fades, cross-fades, dissolves and sudden brightness
changes. That’s where the motion compensation model would usually break
down – whereas with weighted prediction, you can compensate for the
special effects within the coding loop.
With just these five of several MPEG-4 AVC additions – the use of
flexible GOP structures, the partitioning of a 16x16 block in many ways,
the use of highly-accurate quarter-pel motion prediction, the use of
multiple reference pictures, and the use of weighted prediction –
there’s an enormous amount of additional power that can be applied to
video encoding. And that’s where most of the MPEG-4 AVC encoding gains
come from.
Why does your encoder use DSPs/FPGAs instead of ASICs?
Developing on an ASIC, at this early point in the lifecycle of MPEG-4
AVC, just isn’t the right approach. Perhaps it becomes a good idea in
two to three years, when we’ve achieved a significant portion of the
potential of MPEG-4 AVC encoding as a way to cut down on cost and power.
But for now, we think it makes sense to encode on DSPs/FPGAs, for the
simple reason that we can continually improve the video quality of our
MPEG-4 AVC encoder by software download, without asking our customers to
change out their entire encoder.
|
