Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

psnr vs. error energy

0 views
Skip to first unread message

Ryu Chul

unread,
Apr 27, 1996, 3:00:00 AM4/27/96
to

Hi, all.

I am doing some research on motion estimation / compensation.
Recently I'v come up with a new algorithm which finds motion
vectors between frames.

The motion compensated frames after decoding using my algorithm
has higer psnr compared with frames using regular block matching
algorithm.

But the error energy of motion compensated frames before encoding
using my algorithm is higher than the one using BMA.
(before encoding means error energy between reference block and
best-matched block before quantization)

I know that this error term is later added to motion compensated
interframes. So larger error at the encoder means larger error
at the decoder. Right?

Then what did I do wring??

Can anyone tell me what I've missed??

regards,

ryu


Dzung T. Hoang

unread,
Apr 27, 1996, 3:00:00 AM4/27/96
to


Ryu Chul (cr...@photon.poly.edu) wrote:
:
: Hi, all.

Doing motion estimation to minimize prediction error does not guarantee
the best rate-distortion performance, especially at low bit rates. I
have developed motion estimation algorithms for low bit rate coding that
minimize bit rate, a combination of rate and distortion, and a
fast-to-compute heuristic function. The results are demonstrably better
than standard motion estimation techniques in the context of a H.261
coder operating at around 20 kb/s. I get between 1 and 2dB improvement
in PSNR on typical videoconferencing sequences.

One reason for the improvement is that at low bit rates, motion vector
coding consumes an appreciable fraction of the bandwidth (~ 30%).
Reducing the codelength of motion vectors allows more bits to be put
into coding of the prediction residuals, giving better rate-distortion
performance.

Check out the papers I've co-authored, which are available for
downloading from my home page at http://www.cs.duke.edu/~dth, or I can
mail copies to you upon request.

- Dzung Hoang
--
Dept. of Computer Science, Box 90129, Duke University, Durham, NC 27708-0129
Internet: d...@cs.duke.edu
WWW: http://www.cs.duke.edu/~dth/HomePage.html


Stathis VOUKELATOS

unread,
May 1, 1996, 3:00:00 AM5/1/96
to

cr...@photon.poly.edu (Ryu Chul) wrote:

>I am doing some research on motion estimation / compensation.
>Recently I'v come up with a new algorithm which finds motion
>vectors between frames.
>
>The motion compensated frames after decoding using my algorithm
>has higer psnr compared with frames using regular block matching
>algorithm.
>
>But the error energy of motion compensated frames before encoding
>using my algorithm is higher than the one using BMA.
>(before encoding means error energy between reference block and
> best-matched block before quantization)
>
>I know that this error term is later added to motion compensated
>interframes. So larger error at the encoder means larger error
>at the decoder. Right?
>
>Then what did I do wring??
>
>Can anyone tell me what I've missed??

If I understand well what you are saying is that your motion
compensation method produces prediction error frames that have
higher energy than those generated by BMA. This means that BMA
is more successful from a prediction point of view.
However, you say that when you reconstruct the sequence at the
decoder, by adding the quantized error frame to the motion
compensated prediction then you get higher PSNR when your
motion compensated method is used instead of BMA.
One explanation is that although the error frames that are
produced by your method have higher energy they can be coded
more efficiently than the BMA frames, with the coding method
that you are using. For example if you are using transform
coding and your method produces an error frame with more lowpass
characteristics than BMA then it will be more efficiently quantized.
After all, the error that is introduced in the reconstructed sequence
is the quantization error of the error frame and not its energy.

Regards,

--

***********************************************************
Stathis P. Voukelatos
Signal Processing Division
Electronic & Electrical Engineering Dept.
University of Strathclyde
Glasgow G1 1XW

Tel: ++44-141-5524400 ext. 2499
e-mail: sta...@spd.eee.strath.ac.uk
WWW: http://www.spd.eee.strath.ac.uk/~stathis/home.html
***********************************************************


Laurent.Vallette

unread,
May 3, 1996, 3:00:00 AM5/3/96
to

If you are interested in multimedia animation format you can found the
description and the tests of a new one : PPCF (Pixel Proximity Correlation Format)

This format proceeds with 256 colors frames the three strong points of PPCF are
- very fast unpacking rate ( up to 72 images/s format 320x200 on pentium 75)
- compression rate close to GIF
- non destructive algorithm

(this is not a LZW based algorithm)

Address: www.lifl.fr/~vallette
=> choose langage (English or French)
=> go to PPCF link

In order to obtain additional information you can mail me at:
vall...@lifl.fr


Pattabiraman Subramanian

unread,
May 3, 1996, 3:00:00 AM5/3/96
to

On 27 Apr 1996, Ryu Chul wrote:
> I am doing some research on motion estimation / compensation.
> Recently I'v come up with a new algorithm which finds motion
> vectors between frames.
> The motion compensated frames after decoding using my algorithm
> has higer psnr compared with frames using regular block matching
> algorithm.

There is no point in talking, only in terms of distortion. Things have to
be looked at from a rate-distortion point of view if it is to be judged
for optimality. Does your scheme give higher psnr for approximately the
same number of bits per frame used by the regular block matching algorithm?



> But the error energy of motion compensated frames before encoding
> using my algorithm is higher than the one using BMA.
> (before encoding means error energy between reference block and
> best-matched block before quantization)

Do you use the same quantization step size for coding the residual error
for both the schemes ? If you use a smaller quantization step size then
the final psnr of your scheme might be better than regular block matching
although your prediction frame might be poor.



> I know that this error term is later added to motion compensated
> interframes. So larger error at the encoder means larger error
> at the decoder. Right?
> Then what did I do wring??
> Can anyone tell me what I've missed??

Also keep in mind the distortion fuction you minimize. Suppose the
distortion function you use is the SAD (sum of Absolute difference) then
you have to keep in mind that lower distortion does not necessarily mean
greater psnr, bcos psnr use the square of the difference. But if u use
MSE then lower distortion means greater PSNR.

note a + b < c + d does not imply a^2 + b^2 < c^2 + d^2

Hope that helps,

Pattabiraman

____________________________________________________________________________

Pattabiraman Subramanian

# 215, Siegel Hall, 2941, S. Michigan Ave,
Illinois Inst. of Tech, # 101,
Chicago. IL. 60616. Chicago. IL. 60616.
Telephone: (312)-567-3913 Telephone: (312)-326-3023

Email: ra...@kronos.ece.iit.edu or sub...@charlie.acc.iit.edu
URL: http://www.ece.iit.edu/~raman/
_____________________________________________________________________________


0 new messages