Rate-Distortion comparison chart

[koliva]

Hello,

Let say I have a new CODEC and I want to compare my codec with the other current CODECs. As all we know, rate-distortion charts are the most widely used charts for comparison. At the one side of the chart, there is PSNR value and I know how to calculate it. On the other hand, there is Bitrate(bps) values at the other axis of the chart. Could you please tell me how to calculate it. Are they calculating the input data size and the output data size?

Thanks. :)

[Dust Signs]

Hi,

normally you just encode the same sequence at different bit rates (or if your encoder does not support any form of rate control yet: vary the QP) and measure the PSNR achieved with this bit rate. Then plot the PSNR over the bit rate - done.

Dust Signs

[koliva]

Hi,

normally you just encode the same sequence at different bit rates (or if your encoder does not support any form of rate control yet: vary the QP) and measure the PSNR achieved with this bit rate. Then plot the PSNR over the bit rate - done.

Dust Signs

Varying QP will affect on only residual image. Does this changing affect the results that much? Because most of the time I see bitrate varies from 100-500 kbps.

How does the rate control system work in the H.264?

Thanks.

[Dust Signs]

Varying your QP will change your bit rate, although the points in your rate distortion curve will not be equidistant then. For comparison purposes this should not be a problem as long as they cover the same bit rate interval for both encoders with as many points of measurement as meaningful and necessary.
To find out how rate control works refer to Google. It's a very complex topic and there is no standard implementation for it, just more or less similar approaches.

Dust Signs

[koliva]

Varying your QP will change your bit rate, although the points in your rate distortion curve will not be equidistant then. For comparison purposes this should not be a problem as long as they cover the same bit rate interval for both encoders with as many points of measurement as meaningful and necessary.
To find out how rate control works refer to Google. It's a very complex topic and there is no standard implementation for it, just more or less similar approaches.

Dust Signs

Thanks a lot. I think QP will be enough for me.
One more question. In order to measure bit rate, I will start encoding and after one second I will stop program and count how many frames encoded and what is size of the encoded data. Is this the correct way of measuring the bir rate?

[Dust Signs]

No, you should always encode the same number of frames in order to gather reproduceable results. Most encoders have a command line option (or GUI option if there is a GUI) for the maximum number of frames to be encoded. It is also recommended to test one of the VCEG or VQEG test sequences. They are mostly short (some seconds up to half a minute) which would allow for a complete encode.

Dust Signs

[Mr VacBob]

Use SSIM, not PSNR, or else you'll just be misled.

(Why the H.264 forum?)

[Dark Shikari]

Use SSIM, not PSNR, or else you'll just be misled.Even SSIM can be misleading.

Also, encoding at a constant quantizer generally cripples ratecontrol; it's only done in some tests so that encoders can be compared without ratecontrol affecting the results.

[Dust Signs]

Also, encoding at a constant quantizer generally cripples ratecontrol; it's only done in some tests so that encoders can be compared without ratecontrol affecting the results.

Assuming the encoder he is talking about has no form of rate control implemented (maybe an early version or something) QP variation might be the only possibility for him to derive a rate distortion chart. Of course an encoder without rate control would nowadays be useless for a lot of applications but that's a different topic.

Dust Signs