Ateme H.264 HP Beta Test : Bugs & Issues

[Manao]

jfehr : DV often have a yV12 sampling issue, check that out for explanation on the avisynth forum

[couscous]

I have a problem with the ateme decoder/parser
It seems that this is a conflict between ateme decoder/parser and nero digital decoder/parser
When I register ateme decoder/parser and I disabled Nero decoder/parser in MPC, all is OK.
But if I unregister Nero suite, MPC doesn t recognise Ateme decoder/parser anymore nor graphedit (but I can build a graph with Ateme decoder/parser filter and play the clip)
The same thing arrives when I install Nero stuff : I have to register again the ateme decoder/parser... strange...

couscous

[LigH]

That's basically the same as the first issue I posted in this thread...

[IgorC]

the same problem here MPC doesnt recoginize Ateme decoder, but graphedit
plays fine : Ateme Parcer -> Ateme Decoder -> Render

[couscous]

That's basically the same as the first issue I posted in this thread...

ok sorry then :rolleyes:

couscous

Update : Finally no that is not the same issue you talked about...!
Here it's a conflict between filters and not a decoding problem.

[IgorC]

But here problem is persist. I already put low merrit to Nero decoder and even block it and put high merrit to Ateme Decoder but it changes nothing. How to change prior for decoder in Gspot 2.52 b1?
Windows XP SP2

[CyberGuy]

But here problem is persist. I already put low merrit to Nero decoder and even block it and put high merrit to Ateme Decoder but it changes nothing.
Windows XP SP2
Try un-registering the Nero codec.

Un-registers Nero Codec:
REGSVR32.EXE /U "C:\Program Files\Common Files\Ahead\DSFilter\NeVideo.ax"

Registers Nero Codec:
REGSVR32.EXE "C:\Program Files\Common Files\Ahead\DSFilter\NeVideo.ax"

[LigH]

This probably even works from GSpot:

System - List Codecs and Other Filters - (Select desired filter) - Rightclick - ...

[couscous]

Regarding the aspect of interlaced content, it looks like an issue with the video renderer that de-interlaces (badly). The decoder filter does not de-interlace the video, but signals to the renderer that the content is indeed interlaced. It's then up to the renderer to act correctly.
There is an option in the decoder filter to force signalling the interlace content for the MBAFF case.

yep, with my interlaced encoded clip with interlaced,mbaff tool enable, the clip is jerky when it's decoded with ateme decoder/paser but the clip seems to be desentrelaced. When it's decoded with nero decoder/parser, it's not jerky but interlaced.

[IgorC]

This probably even works from GSpot:

System - List Codecs and Other Filters - (Select desired filter) - Rightclick - ...


It was another problem. My windows didn´t accept spacebars or '-' in the name of the *.dll . When the name of *.dll was changed it was registered without problems. Starting to test :). I noticed Ateme Decoder requeires more CPU. On my Cel 2ghz CPU usage 70-95% on resolution 720x304 23.976

[LigH]

The Ateme decoder seems to be very slow on my system, compared to ffdshow 2005-06-11, even when I try to disable deblocking in the Ateme codec. ffdshow is able to play a 1000x540 px video at almost realtime, and - in my opinion - with smoother images (concerning the H.264 special deblocking) than the output of Atemes decoder.

Furthermore, it seems that the decoder forces to use hardware overlay for display, although I prefer the VMR9 renderer in MPC. My graphics card does not support YV12 overlays well, I see some kind of chrominance blocking/shifting in this case. Screenshots by MPC 6.4.8.4 as BMPs do not show this shift, it is only visible on-screen while playing for me. In ffdshow, I disabled YV12 overlay, the next similar mode is YUY2 overlay which works well.


AMD Duron 800 MHz, ABIT KT7-RAID (VIA 686 KT133), 256 MB RAM PC-100 (Durons don't support more on KT133), ELSA Gladiac GeForce2 GTS 32 MB
__

@ IgorC:

In that case, the filename variable at the end of the FOR loop line in your batch file might not have double quotes around it. With double quotes it shall work well even for filenames with spaces:
FOR %%x IN (*.dll *.ax) DO regsvr32 /s "%%x"

[spyder]

I found what appears to me to be a bug in the interlaced output of the decoder filter. I encoded a TFF clip here with xf8x8,interlaced,paff,mbaff at 800kbps 2pass normal quality... When i play this backthere is a horrible jerkiness in the video that can only be corrected by either disabling the "Interlaced material" option or checking the reverse field order one. I can provide the encoded sample if you need it. It's as if the decoder ignores the field order or the encoder specified it wrong.

PS: I'm going out for the night but i am leaving this encoding the same clip with different combinations of interlaced options to see what effect it has.

[ChronoCross]

I don't know what to classify this as so I'll just put it out there. I had the encoding window open and it froze at the same percentage for about 30 mins or so, but when I minimized it and then brought it back up it actually unfroze and continued along it's merry way.

[IgorC]

Source DVD 25 fps, 720x576 resized to 720x544
Codec crashed with the following settings :
encavc.exe -i original.avs -o qu.mp4 -qual extra -rcmode 2pass -br 1500000 -psy 2 -deblock -2 -enhchrp -ref 16 -mvrange +512 -maxb 2 -setef xf8x8,interlaced -mbaff -bref 3

Codec doesn´t crash whithout -setef interlaced

[Manao]

Igorc : your command line is invalid, instead of -setef xf8x8,interlaced -mbaff, you should use -setef xf8x8,interlaced,mbaff. It might only be a bug in the input parameter parser.

ChronoCross : at which priority were you encoding ?

Spyder : can you try with another video renderer ( which if possible doesn't deinterlace )

[LigH]

@ Manao - to Spyder:

I wonder if the Ateme decoder forces a specific renderer (e.g. YV12 Overlay), because with this decoder I see video distorsions I don't see when using ffdshow (see above). To me, it appears as if it was not possible to chose another renderer in the player.
__

@ IgorC:

A "codec" is part of the operating system. You instead are working with a (standalone) "encoder". ;)

[spyder]

@Manao:

I just tried opening the file via avisynth directshowsource and it appears fine there. It must be something in the renderer but it makes little sense to me that it can't simply deinterlace a normal top field first interlaced video. I noticed some strange rendering issues also with VMR9.

Also, just a quality note, it seems mbaff may be the cause of my crappy encodes. I only did one so far with just paff but it's much better. May also depend on the content so I will check some commercials or something tomorrow.

@LigH:

I tried ffdshow(June 11) but it just crashes.

[Manao]

spyder : from our test, MBaff has always been constitantly better than paff ( by a fair margin ). And there's no reason for it to be worse, in fact. So if the quality drop is confirmed, it's a bug

FFDShow ( ffmpeg ) doesn't support interlaced at all ( well, no in fact it might supports i frames, but that's all )

[Blue_MiSfit]

Question - has anyone created or considered creating a simple GUI for us beta testers? some little frontend that will make a bat file to control the encoder? If only I knew some visual basic! :)

[berrinam]

@Blue_MiSfit: A list of options? I don't have the beta.

[LigH]

Considered ... yes. But I'd expect changing command sets during the development, so I wonder if it's really that necessary. Instead, I use a few batch files in the meantime. But maybe tonight or tomorrow... No promises.

[kwtc]

About playing interlaced content.

The Ateme H264 decoder DirectShow filter signals to the downstream filter whether or not the sequence is interlaced, through the VIDEOINFOHEADER2 structure.

Downstream filters (usually a video renderer or a color space converter followed by a video renderer) may ignore this flag or not. The default video renderer will deinterlace, but very badly. For better output, this renderer must be put in video mixing mode (which I don't know how to do with MPC).

The filter has an option to disable usage of this interlaced flag. Open the filter property page and unchek the relevant checkbox. Thus the video renderer will never try to deinterlace.

Otherwise, the filter does not force any particular renderer. It can be connected to anyone that accept YUV 4:2:0 planar. If there is no renderer that accept this format, an intermediate filter must be used to make a conversion to RGB.

Note also that there should be no conflicts between this filter and the Nero ones, apart from the fact that they both can decode H264...

[kwtc]

I have made an other test with an 2000 frames interlaced DV source and the same problem with the control rate is there.


We will look about this one. Thanks for your bug report !

[LigH]

While playing with form designers, thinking about option dependencies ...

"-clref ipred" cannot make sense, IMHO; I won't even think about implementing that into a GUI.

[spyder]

@Manao : regarding the mbaff/paff thing... I haven't made a sample yet using only mbaff so i'm doing that now. I can calc some PSNR or something on each also.

PS: I still don't understand why reversing the field order in the filter properties corrects the playback.

[ChronoCross]

ChronoCross : at which priority were you encoding ?


Normal 1 thread

[babayaga]

Hi guys,
I have a big problem with the "interlaced" tool.
I have made a test with a interlaced clip (Friends) with 500 frames.
I've just compared three encoding settings :
- no "interlaced" tool
- simply "interlaced" tool
- "interlaced" with "paff" and "mbaff" tool
...
As you can see in the second clip, the size is no good (560.39 kb/s instead of 900 kb/s)

As you pointed it out, there is an issue with interlaced or paff and 2 pass rate-control. :mad:
This explains why the final size is not correct. It will be fixed in the next version.

[LigH]

A Taste Of Things To Come


http://www.ligh.de/pics/Atemaker.png


Will take som time. Probably won't be finished ever...

Sorry for the probably crazy layout, but I won't spend too much efforts in it.

What I'm still missing, e.g.:

- What means which psy option? (You are so proud of mode 3 - why?)
- Will it make much sense to specify the CPB?
- Which MV ranges are possible/useful?
- Forgot interlacing... :o Why as bitflags? Does it make sense to e.g. use "-setef interlaced,paff,mbaff", or maybe even "-setef paff,mbaff -clref interlaced"? Can those options be combined, or will one supercede another?

[acidsex]

Ligh, that looks tight. I, for one, thank you for the work you put in to it and hope to have a chance to use it to help speed up my tests by eliminating a lot of the key strokes.

[Andrey]

Finished Die Hard 3 encode, investigating the results :)

1.
FYI: Interesting "percent complete" behavier on my machine when doing first pass
It goes in such a way
1.23%
1.24%
1.25%
1.72%
1.73%
1.74%
2.41%
2.42%
regulary jumping for an about a .5% after some steps...

2.
After going from full to extra quality second pass became 5x faster.

3.
Still do not understand, on what speed of first pass is based.
608x304 clip with br 110000 and length 1000 frames has speed of about 23fps, while 672x336 br 950 and full length was almost twice faster: about 41fps in average...

[Manao]

LigH :
* psy 0 -> no psychovisual heuristics
* psy 1 -> psychovisual heuristics on a frame level
* psy 2 -> psychovisual heuristics on a macroblock level
* psy 3 -> enhanced psychovisual heuristics on a macroblock level

* CBP has its main interest when the CBR rate control is used. It amounts to the VBV buffer in ASP. The larger, the more the bitrate can vary. In ABR / VBR / 2pass, it is also used to define the min / max allowed bitrates.

* though the norm allows unconstrained motion vectors length, level might restrict them to something like 128 or 256 ( which is enough ).

* interlacing : why flags ? well, mbaff and paff can be used together ( and that's not pretty to handle... ), so they need independant flags. Interlaced is more subtile. It can be used alone ( the whole video is encoded as if it were only made of fields ), and it is automatically set by encavc if mbaff or paff is enabled. So, to sum up : interlacied handling has five levels : progressive ( none ), interlaced ( interlaced alone ), paff ( interlaced + paff, or paff alone ), mbaff ( interlaced + mbaff, or... ) and paff + mbaff ( the three of them ).

Apart from that, you forgot 'open gop'

[Mr_Schizo]

Still do not understand, on what speed of first pass is based. 608x304 clip with br 110000 and length 1000 frames has speed of about 23fps, while 672x336 br 950 and full length was almost twice faster: about 41fps in average...
Search in the thread from the first beta test for it, it is explained there. You'll get higher speeds if you're encoding more than 10000 frames cause the encoder uses some special (even patented afaik) techniques.

[LigH]

For CBP, which dimension is usual? 100 KB? 1 MB? 10 MB? Will most probably depend on the frame dimensions and the reference counts, at least.

Interlacing: Well, I guessed so, but had to be sure; so I'll have to restructure my GUI to reflect these dependencies.

Open GOP - sorry, cannot find any option in the documentation regarding the GOP structure, beyond min/max sizes.

[babayaga]

For CBP, which dimension is usual? 100 KB? 1 MB? 10 MB? Will most probably depend on the frame dimensions and the reference counts, at least.

It's very dependant on the applications.
Some use 224kB wich is a standard value in MPEG-2 and very customary AVC/H.264. The default value in encavc is 1MB (it will hopefully default back to 224kB in the next version).

Some express it in seconds (because this buffer size is introduces a delay for the decoder), thus depending on the bitrate. Standard values are 1s to 5s.

[LigH]

Well - so it seems to be best to leave this decision to the encoder, unless you know why not. ;)
__

Uploaded an updated version to the FTP incoming. I hope bobololo will then make it available for the testers. ^ Screenhot updated on my server, too.

[bond]

i stumbled over the foreman-ateme-lossless.mp4 encode and noticed that it shows wierd colors when played with latest ffmpeg

when i extracted the stream to raw and ran mpeg4ip's h264_parse over it i only got the SPS, PPS and first SEI

the used encoder version was 1.2.0.9. is that a known bug?

[IgorC]

Interlace mbaff doesn't work at all here(enocoder 1.2.0.14) . Ouput is horrible and doesn't matter what settings of decoder were used. I also try ffdshow (without deinterlace) to decode it to see if it's problem of decoder. The same result. Waiting for a new beta. Gui for CLI is welcome for comfortable and fast test :)

[CyberGuy]

For CBP, which dimension is usual? 100 KB? 1 MB? 10 MB? Will most probably depend on the frame dimensions and the reference counts, at least.

Interlacing: Well, I guessed so, but had to be sure; so I'll have to restructure my GUI to reflect these dependencies.

Open GOP - sorry, cannot find any option in the documentation regarding the GOP structure, beyond min/max sizes.
-opengop

[CyberGuy]

I had doing some peeking and found some other command line options. Some I know like 1st, 2nd, subpart, pcm, -maxbr, and –opengop, but what is –spmvp, -log, debug and analysis used for?

[acidsex]

Ok, ran some more tests using some of the Lynda.com training files I have. The source files are 880x660 5fps, encoded with Sorenson 3 Pro. I loaded them into Vegas and output to HuffYUV using the same resolution.

I used the following command :" encavc.exe -i clip.avs -o clip.mp4 -qual extra -rcmode 2pass -br 1100000 -psy 2 -deblock -2 -adaptdbk"

Quality is quite well but file size is almost 2 1/2 times the original file size. If I lessen the bit rate, I do get a smaller file size but poor quality when compared to the source. Which leads to my big question. Is Sorenson 3 Pro that much more efficient at encoding video than Ateme's AVC? I would think not but right now, the tests I've done show differently.

Using Sorenson 3 Pro, I can archive 6-8 hours of video at 5fps on one 700MB CD. In order for me to encode the same amount of video using H.264 AVC at 5fps it takes almost 3.8GB to store the same video where text is legibile and not fuzzy.

Can anyone help explain this or is this just a limitation of the H.264 codec? I can make a copy of the source file I am using available to test this if needed.

[Manao]

IgorC : ffdshow CAN'T decode mbaff. If you want to test it, try with nero's decoder.

acidsex : your source is already compressed a lot, you can't compare the codec like that

CyberGuy : they are mostly useless remnant of the first beta, or debugging options. debug is a quality level, analysis isn't a command line parameter, and i think spmvp means spatial mv prediction in bframe ( instead of temporal ).

[*.mp4 guy]

I'm not sure if this is an error, but it looks a little off that the bitrate for the first pass is so low.
http://x2.putfile.com/6/16902095997.png

[ChronoCross]

seems pretty much insignificant in terms of total accuracy. it wouldn't even have a noticeable quality loss/gain. now if it was off by like 30-100 then I would be worried as that can produce a pretty severe difference in the quality of it.

[Manao]

*.mp4 guy : the first bitrate reported is the bitrate achieved on the last 32 frames of your clip, while the last one is the overall bitrate of the second pass. Encavc doesn't give you the bitrate of the first pass.

[plonk420]

my encode missed the bitrate by a LONG shot
http://plonkmedia.temp.powweb.com/bigmiss.png
on the other hand it IS interlaced...

[Manao]

When compressing interlaced content, since paff and interlaced are buggy, if you want to hit the bitrate xxx, you have to ask for the bitrate xxx * 3 / 2 .

[plonk420]

When compressing interlaced content, since paff and interlaced are buggy, if you want to hit the bitrate xxx, you have to ask for the bitrate xxx * 3 / 2 .

hrm, a) did i miss that somewhere in the thread? :eek: and b) will this likely be fixed soon or should i just count on using bitrate * 3 / 2 for a while?

[*.mp4 guy]

*.mp4 guy : the first bitrate reported is the bitrate achieved on the last 32 frames of your clip, while the last one is the overall bitrate of the second pass. Encavc doesn't give you the bitrate of the first pass.
Ahh, that explains it ;) I was a bit baffled by that first number for a while.

[Manao]

plonk420 : it will surely be solved for the next beta release. But have a look at the ratio bitrate wanted / bitrate obtained in this thread when paff / interlaced was used, it will always be 3 / 2 ( 900 -> 560, 1600 -> 1068, and i tried 1000 which gave we 666.68 and hinted me more obviously what had to be done to 'solve' temporarily the issue )

[IgorC]

IgorC : ffdshow CAN'T decode mbaff. If you want to test it, try with nero's decoder.


Yes, I tried to decoded it with Ateme decoder 2.0.1.0 ( with Interlaced material on and off) with this settings

encavc 1.2.0.14

encavc.exe -i corto.avs -o qu.mp4 -qual extra -rcmode 2pass -br 1500000 -psy 2 -deblock -2 -enhchrp -ref 16 -mvrange +512 -maxb 2 -setef xf8x8,interlaced,mbaff -bref 3

When I tried to do screenshots the image of output was good. But during playback all objects have not normal curve borders.

Windows XP SP2. Celeron 2ghz. Source : DVD PAL 720x576 (4:3) resized to 720x544