2pass worse SSIM-Marks than 1pass

[fankler]

Hello,

I'm trying to find out how much quality I can win if I use 2pass instead of 1pass in wm9. But if I compare the two encoded frames with Lefungus' SSIM there are some clips that are worse in 2pass mode. Especially in a scene from Alien vs. Predator (uncompressed source) in which two persons running in front of a firewall.

Another point I can't explain to myself is: I reduced the horizontal resolution to 480 pixel (2/3 PAL). If I encode this low-res file and compare it to its full-res equivalent which is encoded at the same bitrate, the one in full-res gets better marks. (DVB-S MPEG-2 TS source)

Video-Bitrates were between 1152kbps and 2688kbps. SSIM measurement was made running avs in Windows Media Player Playlist.

Has anybody an idea how this can be?

Thanks

fankler

[MfA]

Reducing resolution before compression is a very crude trick, a good codec will always do better by encoding the original resolution (not to say there aren't a lot of sucky codecs in this regards though).

[bond]

are these scenes also visually worse? after all what your eyes are seeing counts, not some ssim value

[IgorC]

did you check final size of 1-pass and 2-pass videos?

[Kopernikus]

Another point I can't explain to myself is: I reduced the horizontal resolution to 480 pixel (2/3 PAL). If I encode this low-res file and compare it to its full-res equivalent which is encoded at the same bitrate, the one in full-res gets better marks. (DVB-S MPEG-2 TS source)


How did you compare two files with different resolutions?

[fankler]

How did you compare two files with different resolutions?

I compared the full-res encoded clip to the full-res original clip and the low-res encoded clip to the low-res original one. If I spend the same bitrate to fewer pixel, this should mean more bit per pixel -> better quality. Isn't SSIM proposed to "see" this?

did you check final size of 1-pass and 2-pass videos?

2pass is slightly bigger. This is another point I have to figure out in the next days. I've checked three times whether I made some wrong configs but they are all the same but the checkbox "use 2pass".

are these scenes also visually worse? after all what your eyes are seeing counts, not some ssim value

I knew this would be one of the first answers. But I want to know is how much I can rely on SSIM results at least as a indicator whether quality is better or not (not even how much it is better). I was shure 2pass always produces better quality, so I didn't check the quality visually by now (I will do so this afternoon - I hope).

[zambelli]

The likely problem is that you're generating the SSIM average index based on the entire movie. I never saw AvP, but I'm guessing that like most movies, it's dominated by low motion scenes with occasional bursts of high motion scenes. Therefore, the average SSIM index for the entire movie is going to be weighed heavily in favor of the low motion scenes.
In CBR mode, the entire movie gets encoded with equal bitrate which means that the low motion scenes will all look about the same quality, but the high motion stuff will look like crap.
If you were to encode the entire movie in 2-pass VBR mode, the high motion scenes would look better than in the CBR example, but many of the low motion scenes would look worse because their bitrate was now likely to be distributed elsewhere. However, because the low motion scenes outnumber the high motion scenes, the overall SSIM index would actually sink lower than for CBR.

Well, that's just my theory anyway. You can try this for experiment: instead of generating SSIM for the whole movie, pick two scenes from the encode, one low motion and one high motion. Calculate the SSIM for each scene separately, then compare the CBR and VBR numbers. I predict that the low motion SSIM will slightly favor CBR, but the high motion SSIM will show a huge improvement in VBR.