AVC Decoder Chroma Quality Comparision

[CruNcher]

Test System:
CPU: Athlon 64 3000+ @ 2.4 Ghz (Venice)
GFX: Nvidia 7600 GS PCI-E / Forceware 94.24
OS: Windows XP SP2
Player: MPC Home Cinema 1.0.9.0
Renderer: VMR9 Windowed (TV Scale)
Sample: Philips Colors of Miami (1080i Mpeg-2) -> x.264 (720p) L4.1 High Profile

Software Decoder:

Ateme 2.2.1.0
CoreAVC 1.5
Cyberlink 2.0.0.1904
Elecard/Mainconcept 1.0.41.60818
ffdshow tryouts 1405 (YV12)
Nero 4.11.3.1 (Ateme/SDK based ?)
Sonic Cinemaster 4.2 (Elecard/Mainconcept SDK based)


Hardware Decoder: PureVideo 1

Cyberlink 2.0.0.1904 (has problems with certain x.264 (mp4/mkv) combinations (still analyzing see http://forum.doom9.org/showthread.php?t=127376
http://forum.doom9.org/showthread.php?p=1022197 and http://forum.doom9.org/showthread.php?t=124945)

Elecard/Mainconcept 1.0.41.60818 (newer versions 2.0/3.0 available run only in Software mode with my Test System no answer ever recived from Elecard/Mainconcept about this issue see http://www.elecard.com/forum/viewtopic.php?t=1793)

Nero 4.11.3.1 (it's claimed to be a Hardware Decoder Mpeg-2/Mpeg-4 AVC/VC-1? but i never get Hardware Decoding working with my System neither in Showtime nor any other Directshow Player see http://forum.doom9.org/showthread.php?t=113516)


Results:

Ateme 2.2.1.0 (Fastest configuration enabled)
http://www.directupload.com/files/vhxznem2mudgwzmu4zdo.png

CoreAVC 1.5 (TV Scale) (YV12) (Chroma Quality reported for 1.1.0.6 fixed)
http://www.directupload.com/files/402eeyn212zzmenmjyyg.png

Cyberlink (Hardware PV1)
http://www.directupload.com/files/mlwoknin0mymmtgvmmjt.png

Cyberlink (Software)
http://www.directupload.com/files/mbjtgqzt5myxzyjmmqx0.png

Elecard/Mainconcept (Hardware)
http://www.directupload.com/files/3mjdt5diazyzoiaydxdo.png

Elecard/Mainconcept (Software)
http://www.directupload.com/files/wiinvi2gozywxytozfhn.png

Elecard/Mainconcept (Software HQ Upsampling ON)
http://www.directupload.com/files/zwwntge4jmm42ty5ynnm.png

ffdshow (YV12)
http://www.directupload.com/files/ndyjnu2dg0o2qmy230ez.png

ffdshow (YV12->RGB32 High Quality)
http://www.directupload.com/files/4tqynzjnmzmntnzzrkkt.png

Nero (Software)
http://www.directupload.com/files/ogw0m1ym31kjwmnivzgm.png

Sonic Cinemaster
http://www.directupload.com/files/dhqm2yqy2wzdnjmizjkz.png

PS: Don't ask me why some Decoders show this slight luma difference to the others ;)

Coming Soon (AVC Decoder Benchmark,Next Generation Codec Detail Preservation Test includes Noise/Grain,VC-1 Intermediate test)

[foxyshadis]

CoreAVC's internal RGB output is just as bad as it always was, though; I still have to pipe it through ffdshow to get smooth reds.

[CruNcher]

My results show the same quality as ffdshow now and virtualy every other decoder (a year ago CoreAVC looked like the Elecard/Mainconcept, Sonic Cinemaster results) except Elecard/Mainconcept <- High Quality Upsampling enabled has no effect and when useing Hardware it's allready active via PV1 in this example same goes for Cyberlink no Software upsampling only Hardware.

Maybe i should have said in wich area you see the problems the girl in red ,bra,hat and the background 8 on the ship, with a good Chroma upsampling it would be smooth a bad one shows pixelated edges :)

also you can see that it makes no reall difference wich Hardware Decoder (PV1) you use the results (Quality) will be identical same API same results compared to the software decoders (wich are also identical except sonic cinemaster and mainconcep/elecard because of the missing upsampling) you can also see the output for the Hardware decoded frames is smaller i think thats coming from the NV12 converting and upsampling seems to be a little different to the YV12 the Software Decoders use.

[Theliel]

very interesting

You are right, Cyberlink hadware give better and smooth edges than cyberlink software.

Group 1- Elercard Hardware equal Cyberlink Hardware
Group 2- CoreAVC, Ateme, Nero, FFdshow YV12 and FFdshow YV12->RGB32 are equal bit to bit.
Group 3- Elecard HQ
Group 4- Cyberlink software
Group 5- Sonic and Elecard software are equal bit to bit too.

The best IQ are for group 1. Group 2 are very very closed.
Group 3 and 4 are very similar, sometimes IQ in 3 are better sometimes are 4. Group 3/4 are noticeable worse group 1 in edges.
The worst results are for group 5, where edges are much more noticeable than other groups.

[wozio]

Wait a second. Doesn't almost every video coded (compressed) at nv12/yv12 and decoders usually outputs that colorspace? In that case chroma upsampling is up to renderer not decoder. Of course when decoder outputs yuy2 or rgb then it must do it by himself, then decoder quality matters.

Or maybe I undestand it wrong?

You should add detailed description of format outputted by each decoder.

[Soulhunter]

Right, even if a renderer is able to upsample YV12 -> RGB/YUV correctly, it cant do its job if the decoder feeds it with wrong upsampled YV12 -> RGB/YUV frames... Anyway, its good to see that all the nagging about wrong chroma upsampling had a effect, at least on some devs! =D


Bye

[wozio]

Yes but such comparision doesn't make any sense when for example comparing coreavc outputting nv12 and ffdshow outputting nv12, result will be identical at least on chroma upsampling subject. The same goes for all hardware decoders, all of them internally use nv12 color space and all color conversion (and chroma upsampling) is up to renderer (or driver to be more precise).

Decoders outputting yuy2 or rgb (any except forced ffdshow?) are other story but then it would be great to have detailed description of output media type.