LAV Filters - DirectShow Media Splitter and Decoders [Archive] - Page 125

madshi

15th October 2011, 12:53

madshi, again, the problem is not decoding matrix / primaries. I've posted the sample. You write trivial avisynth script
You didn't say you were testing with an RGB image source, or at least I didn't notice if you did. I thought you were talking about testing YCbCr video all the time. Of course if you feed madVR with an RGB image source, decoding matrix and primaries should have no affect.

It's not a "bug" it just looks like madVR doesn't do true scaling but takes the same shortcut as almost everyone else (ironically, it claims to take "no shortcuts").
madVR does true scaling without any shortcuts. However, madVR currently scales in gamma corrected light. If I modify madVR to scale in linear light, I get similar results to yours:

http://madshi.net/gammaLight.png
http://madshi.net/linearLight.png

These are madVR screenshots, using exactly the same algorithms (SoftCubic for Chroma, Lanczos for Luma). The only difference is gamma corrected light vs. linear light.

It seems that with your specific test image, linear light produces better results. That is caused by the nature of the images, which contain high contrast dithering patterns. When using linear interpolators in gamma corrected light, the blending of the high contrast dithering patterns produces visible errors. Switching to linear light seems to fix this problem. FWIW, this problem seems to be mostly limited to such special images with high contrast dithering patterns. With real life movie content the difference between gamma and linear light scaling seems to be much smaller.

What I can say is that all the current hardware chips out there, even the highest quality ones like HQV Realta and Gennum VXP, are scaling in gamma corrected light, too. I've been planning to add an option to madVR to scale in linear light for a while, just haven't had time to implement it yet.

VipZ

15th October 2011, 13:03

That file also crashes PotPlayer when doing few fast seeks. No issue if PotPlayer's VP8 decoder is used.

Nev, when opening an DVD IFO file in MPC with LAV Splitter in use, the file duration is only that of the 1st vob, but the entire movie is there and seekable. With MPC's MPEG splitter it has the correct duration. This happens on 0.30 to 0.37 in my tests and don't have anything pre 0.30 to test.

nand chan

15th October 2011, 13:05

You didn't say you were testing with an RGB image source, or at least I didn't notice if you did. I thought you were talking about testing YCbCr video all the time. Of course if you feed madVR with an RGB image source, decoding matrix and primaries should have no affect.

madVR does true scaling without any shortcuts. However, madVR currently scales in gamma corrected light. If I modify madVR to scale in linear light, I get similar results to yours:

http://madshi.net/gammaLight.png
http://madshi.net/linearLight.png

These are madVR screenshots, using exactly the same algorithms (SoftCubic for Chroma, Lanczos for Luma). The only difference is gamma corrected light vs. linear light.

It seems that with your specific test image, linear light produces better results. That is caused by the nature of the images, which contain high contrast dithering patterns. When using linear interpolators in gamma corrected light, the blending of the high contrast dithering patterns produces visible errors. Switching to linear light seems to fix this problem. FWIW, this problem seems to be mostly limited to such special images with high contrast dithering patterns. With real life movie content the difference between gamma and linear light scaling seems to be much smaller.

What I can say is that all the current hardware chips out there, even the highest quality ones like HQV Realta and Gennum VXP, are scaling in gamma corrected light, too. I've been planning to add an option to madVR to scale in linear light for a while, just haven't had time to implement it yet.

Have you considered scaling in XYZ space? What would that even look like?

nevcairiel

15th October 2011, 13:28

Same here. LAV Splitter + LAV Video + mplayer2
when playing with avplay and seek, sometimes it shows a yellow "[vp8 @ 021621a0] Discarding interframe without a prior keyframe!"

You're saying it happens with mplayer2 as well (which isnt even a DirectShow player)?

I could reproduce it now, it requires quite alot of seeking. Apparently it really doesn't like starting decoding without a key frame. For some reason it dead-locks then and doesn't return from the decode call. Looks like a ffmpeg bug to me (related to multi-threading, i guess)

Once i start working on seeking improvements in the splitter, this would be avoided.

fastplayer

15th October 2011, 13:33

For some reason it dead-locks then and doesn't return from the decode call. Looks like a ffmpeg bug to me (related to multi-threading, i guess)
Bingo! I disabled multi-threading and seeking works as usual.

madshi

15th October 2011, 13:33

Have you considered scaling in XYZ space? What would that even look like?
Isn't XYZ linear light, too? I don't think there would be any visible difference. But I'm kinda guessing here...

nevcairiel

15th October 2011, 13:56

Bingo! I disabled multi-threading and seeking works as usual.

I applied a fix so that it'll always wait for a keyframe before it actually sends data to the decoder. This stop the freezes.

I should really get onto those seeking improvements in LAV Splitter.

Thunderbolt8

15th October 2011, 13:59

nand chan

15th October 2011, 13:59

Isn't XYZ linear light, too? I don't think there would be any visible difference. But I'm kinda guessing here...

XYZ is linear intensity just like linear RGB, but the difference is that XYZ is also linear chromaticity. In other words, X Y and Z have no special gamut, relation or proportions attached, so if you interpolate in XYZ space it'll probably be like spreading the raw light in the real world, eg. by using a lense to stretch.

Whether you can see a difference or not would be interesting to check. I'd test it myself but I have no idea how to implement scaling algorithms (eg. bicubic)

nevcairiel

15th October 2011, 14:01

in case of dtsma audio tracks, which information does lavsplitter actually diplay? got a 7.1 dtsma track, but LAV splitter displays "dts-hd ma, 48000 Hz, 5.1, 1536 kb/s"

so at the beginning it says dtsma, but then lists the properties of the core track. what about the rest of the dtsma info and why the mixing of information with the core track properties?

Its rather easy to figure out what kind of profile the DTS uses, however figuring out what is actually inside a HD profile would require quite alot more parsing, and since its only cosmetical, seems unnecessary effort.
So, it will show what DTS profile it is, but it will only show the other infos from the DTS core.

madshi

15th October 2011, 14:11

XYZ is linear intensity just like linear RGB, but the difference is that XYZ is also linear chromaticity. In other words, X Y and Z have no special gamut, relation or proportions attached, so if you interpolate in XYZ space it'll probably be like spreading the raw light in the real world, eg. by using a lense to stretch.

Whether you can see a difference or not would be interesting to check. I'd test it myself but I have no idea how to implement scaling algorithms (eg. bicubic)
Well, if you PM me with code how to convert to/from XYZ, I can give it a try. We should move that to PM, since it's OT here.

nevcairiel

15th October 2011, 14:14

Hi...

Did you find a moment to have a look at that, nev?

Thanks again!

I looked at the file now, and it seems like alot of the frames are just marked as progressive for some reason. Either they were encoded as weaved fields into a progressive image, or something is fishy with the detection.

If you're using the NVIDIA decoder, you can flag "Force Deinterlacing" to make it process the progressive frames as well. If you're using the software decoder + EVR for deinterlacing, i'm afraid there is no solution right now, but i do have some more plans to be able to force deinterlacing there as well.

fastplayer

15th October 2011, 14:28

I applied a fix so that it'll always wait for a keyframe before it actually sends data to the decoder. This stop the freezes.
Thanks! :)
I should really get onto those seeking improvements in LAV Splitter.
Naive question: When you say "improvements", does that mean that ffmpeg's seeking functionality (assuming here there is one) is not good enough?

iSunrise

15th October 2011, 14:40

Downmixing to stereo is a rather easy task. I've only really not done it yet because i kind of wanted to do some more mixing options.

How many people need Stereo downmixing in contrast to how many people "want" 5.1/7.1 upmixing?
Downmixing to stereo would be a killer feature for me. I already use your filters for everything content-wise and downmixing to stereo would make it just perfect. I could finally enjoy decoded DTS-MA without having to insert ffdshow audio into my playback-chain. Upmixing is not really relevant for me (my receiver does a pretty good job at that), but for others, possibly.

nevcairiel

15th October 2011, 15:00

Naive question: When you say "improvements", does that mean that ffmpeg's seeking functionality (assuming here there is one) is not good enough?

No, its seeking functionality sucks. It can seek just fine, however it just seeks to the time you ask it (well, near the time anyway), it doesn't seek to keyframes - for playback thats a required feature.

roytam1

15th October 2011, 15:59

You're saying it happens with mplayer2 as well (which isnt even a DirectShow player)?

I could reproduce it now, it requires quite alot of seeking. Apparently it really doesn't like starting decoding without a key frame. For some reason it dead-locks then and doesn't return from the decode call. Looks like a ffmpeg bug to me (related to multi-threading, i guess)

Once i start working on seeking improvements in the splitter, this would be avoided.

I mean the wmp 6.4 which the executable name is mplayer2.exe.:p

No, its seeking functionality sucks. It can seek just fine, however it just seeks to the time you ask it (well, near the time anyway), it doesn't seek to keyframes - for playback thats a required feature.

Maybe we can did the same thing like CODEC_ID_MPEG2VIDEO and CODEC_ID_VC1 for them in the moment?

nand chan

15th October 2011, 19:42

madVR does true scaling without any shortcuts. However, madVR currently scales in gamma corrected light. If I modify madVR to scale in linear light, I get similar results to yours:

http://madshi.net/gammaLight.png
http://madshi.net/linearLight.png

These are madVR screenshots, using exactly the same algorithms (SoftCubic for Chroma, Lanczos for Luma). The only difference is gamma corrected light vs. linear light.

I'm just going to mention here that it's possible to upscale in linear light using EVR-CP as well (add a pre-resize shader to decode to linear light, and a post-resize shader to encode back to gamma light).

Here's what it looks like, as upscaled using Catmull-Rom spline6: http://img828.imageshack.us/img828/311/upscalelinear.png

Edit: Disabled LittleCMS and added a 16-235 -> 0-255 shader for better comparison with madshi's screenshot (which seemingly mistakes it as limited range instead of full range)

http://img690.imageshack.us/img690/311/upscalelinear.png (here is a result close to madVR's, however this is incorrect as you can easily check with any image editor program)

nevcairiel

15th October 2011, 19:50

added a 16-235 -> 0-255 shader for better comparison with madVR. (I think madVR does this incorrectly)

madVR assumes all RGB sources are full-range, unless it has data to suggest otherwise.
But didn't you guys want to move to PM or some other thread, or some place? :p

SEt

16th October 2011, 00:27

nevcairiel, just let me post one more replay - the issue is clear now ;)

You didn't say you were testing with an RGB image source, or at least I didn't notice if you did. I thought you were talking about testing YCbCr video all the time. Of course if you feed madVR with an RGB image source, decoding matrix and primaries should have no affect.
Well, being png (and not something like ebmp) I thought it's obvious.

madVR does true scaling without any shortcuts. However, madVR currently scales in gamma corrected light. If I modify madVR to scale in linear light, I get similar results to yours
Seriously, how can you name it "true scaling without any shortcuts" when scaling gamma corrected image (99.9...% real world video) disregarding gamma?

It seems that with your specific test image, linear light produces better results. That is caused by the nature of the images, which contain high contrast dithering patterns. When using linear interpolators in gamma corrected light, the blending of the high contrast dithering patterns produces visible errors. Switching to linear light seems to fix this problem. FWIW, this problem seems to be mostly limited to such special images with high contrast dithering patterns. With real life movie content the difference between gamma and linear light scaling seems to be much smaller.
I've written renderer than can satisfy myself for images (manga and photo) first, then reused it for video. For video effect is less pronounced because it's usually more blurry. Still, I'd say correct gamma processing is way closer to real world than effects shown by madTestPatternSource :rolleyes:

What I can say is that all the current hardware chips out there, even the highest quality ones like HQV Realta and Gennum VXP, are scaling in gamma corrected light, too. I've been planning to add an option to madVR to scale in linear light for a while, just haven't had time to implement it yet.
I trust my eyes, not useless (and often false) "highest quality" claims. Scaling photos disregarding gamma produce obviously wrong results.

BTW, same and even more problems apply to chroma scaling. So no way trivial cubic/lanczos/... can be claimed "high quality". Regrettably, here I don't have satisfactory solution myself.

asasadad_1

16th October 2011, 02:26

here (http://uploading.com/files/ca79ab5d/dts.wav/) is a dts wav sample that LAV Splitter+LAV Audio Decoder output noise,dcadec.exe(libdca-0.0.5) works well.

madshi

16th October 2011, 08:39

My last post on this topic in this thread. Sorry for OT.

Well, being png (and not something like ebmp) I thought it's obvious.
The main interest area of Doom9 is YCbCr video. Still practically every single screenshot posted in these forums is either PNG or JPG. I don't even know what "ebmp" is. So you posting a PNG screenshot does not in any way indicate that you were talking about scaling an RGB image source. Anyway, doesn't matter. It was a misunderstanding, cleared up now.

Seriously, how can you name it "true scaling without any shortcuts" when scaling gamma corrected image (99.9...% real world video) disregarding gamma?
Linear light scaling is not commonly used in video processing. To my best knowledge not a single consumer electronics device does linear light scaling for video content. I consider scaling in linear light a nice "extra feature", which has been on my to do list for a long time. madVR has not reached version 1.0 yet, so you can't expect it to be feature complete right now. Scaling in linear light also has disadvantages, e.g. the ringing caused by negative lobes (e.g. when using Lanczos) is increased. So the ultimate solution should be a ringing-free algorithm (e.g. this (http://forum.doom9.org/showthread.php?t=145358)), using linear light. But just because I haven't had the time to implement such a solution yet doesn't mean I've been using shortcuts. I've just not had the time to implement all planned features yet. There is no shortcut in any of the code I've written so far. I've just not written all planned code yet. If you want to complain about missing features, wait for madVR v1.0 to be released. If then something is still missing, you have an actual reason to complain.

I've written renderer than can satisfy myself for images (manga and photo) first, then reused it for video. For video effect is less pronounced because it's usually more blurry. Still, I'd say correct gamma processing is way closer to real world than effects shown by madTestPatternSource
You seem to be coming from the image processing area. Priorities are different. Maybe for image processing linear light scaling is critical, I don't know that. But for video processing I consider banding artifacts to be much more annoying (at least to my eyes), especially if the banding changes from frame to frame.

BTW, same and even more problems apply to chroma scaling. So no way trivial cubic/lanczos/... can be claimed "high quality". Regrettably, here I don't have satisfactory solution myself.
I already have some clever algorithms for chroma upsampling in mind. But again: Just because I've not implemented them yet doesn't mean that I'm using a shortcut right now. BTW, if you don't know this article yet, you may want to look at this:

http://www.glennchan.info/articles/technical/chroma/chroma1.htm

My algorithm plans are different, but maybe that article will give you some ideas.

nevcairiel

16th October 2011, 09:07

here (http://uploading.com/files/ca79ab5d/dts.wav/) is a dts wav sample that LAV Splitter+LAV Audio Decoder output noise,dcadec.exe(libdca-0.0.5) works well.

That file is just off. It has around 60kB of null bytes before the actual data starts. And even once it starts, it generates alot of warnings in the decoder.

Anyhow, i slightly increased the probing buffer size, so that it will actually notice that its a DTS file.

asasadad_1

16th October 2011, 09:38

That file is just off. It has around 60kB of null bytes before the actual data starts. And even once it starts, it generates alot of warnings in the decoder.

Anyhow, i slightly increased the probing buffer size, so that it will actually notice that its a DTS file.

confirmed,thanks for your robust filters.
:thanks:

nevcairiel

16th October 2011, 17:40

Hi guys,

i'm finally working on the new subtitle selection logic, and i could use some feedback on how the default "simple" mode works.

First, let me explain the changes to the settings.
Basically, in "simple" mode, there will be the old input box where you just enter languages, and in addition there will still be the dropdown to configure "No subtitles", "Default/Forced subtitles", "Always subtitles", "Advanced mode".
The "Only with matching language" checkbox is gone, as most people didn't know what it did anyway.

- "No subtitles" is perfectly clear, never load any.
- "Default/Forced subtitles" is still pretty clear. It'll only load subtitles marked as default, or forced. If the user configures languages, it'll prefer those, but if there is no language match, it'll still fall back to the first forced or default track (forced takes precedence).

- "Always subtitles" is where my problems start.
The first part is easy. If there are languages configured, look for a stream matching those languages, and use that one.
Now comes the hard part. What should happen if no language match was found? Or no languages were configured?
Right now, i would look for a track marked default/forced, use that one if found, but if not, show nothing.

I'm not a big subtitle user, so i really cannot comment on whats more common to use for users.

Right now, this seems like a good default. If someone wants more, they can always go into advanced mode and enter "*:*" as a rule, allowing matching of any subtitle stream. :p
I really haven't tested much how Haali works, maybe i should get on that.

Anyway, implementation is coming along nicely, it won't be long until i can show you a version to play with.

SamuriHL

16th October 2011, 17:47

VERY kuel. I know my preference for how things work is probably different than most people. I follow the blu-ray model. Which is this:

-ALWAYS play forced subtitles that match the language of the audio track I've selected
-ONLY play "default" when subtitles are "turned on".

I realize this doesn't make sense to most people, but, that is in fact what blu-ray players do with an actual disc.

So to get back to your new options, I'd have "default/forced" on by, um, default. BANG :D

fastplayer

16th October 2011, 17:49

If no languages are configured, show subtitles always is selected, and no default/forced track is found, then just use the first available subtitle.
"Always subtitles" implies that at least something subtitle-like is shown.

SamuriHL

16th October 2011, 17:49

P.S. I had a problem with forced subs showing off the Star Wars blu-rays...I didn't get foreign language translations in Attack of the Clones yesterday. That's playing off the disc, not in an MKV.

nevcairiel

16th October 2011, 17:50

Do Blu-ray players actually let you configure a preference, like you configure subtitles to be in french, and it'll always load french subtitles? The question then is, what happens if there are no french subtitles? :)
I thought you always had to go through the BD menu on most players.

P.S. I had a problem with forced subs showing off the Star Wars blu-rays...I didn't get foreign language translations in Attack of the Clones yesterday. That's playing off the disc, not in an MKV.
Those can't be auto-detected, they are in a separate stream. Its the second-to-last english track, iirc.

SamuriHL

16th October 2011, 17:52

No. Blu-ray players do not. By default, they give you subs in the language you've chosen for audio. The menus can override the behavior, though, which is currently impossible to deal with in open source/freeware solutions. What I mean is, if I select a foreign audio track on a lot of my martial arts movies, I get english subs by default. I'm pretty sure that's controlled by the menu.

(Well, actually, if I set english subs for preference, I suppose when I pick a foreign audio track using LAV filters, it should figure out what I want)

nevcairiel

16th October 2011, 17:56

(Well, actually, if I set english subs for preference, I suppose when I pick a foreign audio track using LAV filters, it should figure out what I want)

This is all about the initial selection, if you manually change the audio stream, the subtitles will not change (well, with one exception, if its a BD, and the automatic forced stream is active, then that will change to the current audio language, if a sub stream matches that language)

SamuriHL

16th October 2011, 17:58

So to spell out EXACTLY how I want it to work in my setup:

Assuming english audio:
1) If english forced subs are detected, play them
2) If no forced subs - show no subs until I turn on subs, then show default english sub track
3) Otherwise, show no subs

Non-english audio:
1) If english forced subs are detected, play them
2) If no forced subs - show no subs until I turn on subs, then show default english sub track
3) Otherwise, quickly teach me the foreign language :D

So basically, if I set subs to default english, I always want forced subs, and ONLY to show default subs if I explicitly turn subs on in my player. Does this make sense? Tell me how to do that with your new model and I'm good with whatever else you do. :) Thanks!!

SamuriHL

16th October 2011, 17:59

This is all about the initial selection, if you manually change the audio stream, the subtitles will not change (well, with one exception, if its a BD, and the automatic forced stream is active, then that will change to the current audio language, if a sub stream matches that language)

I understand. See my spelled out requirements. Basically my first requirement is all that matters to me because by default, I don't want subs unless they're forced. I will deal with turning subs on if it's a foreign language.

nevcairiel

16th October 2011, 18:00

"Turning subs on in the player" does mean the player needs to select a different track, however.
If you have a player smart enough to control stream selection, this is all null and void, i suppose (like MC16, it overrides any default choices anyway).

There is also always the advanced mode, which can probably do anything you guys can come up with - and if not, i will make it so. :p

I just want to come up with some smart "default" behaviour - i will probably just quickly check how Haali behaves, and try to mirror that as a default so people get what they expect. :p

SamuriHL

16th October 2011, 18:02

"Turning subs on in the player" does mean the player needs to select a different track, however.
If you have a player smart enough to control stream selection, this is all null and void, i suppose (like MC16, it overrides any default choices anyway).

There is also always the advanced mode, which can probably do anything you guys can come up with - and if not, i will make it so. :p

Ok, let me boil this down to something that hopefully makes sense. :D I only ever want "automatic subs" in the case of forced subs being found for my preferred language. I NEVER want to see default subs until I explicitly decide I want them. Does that help? :)

SamuriHL

16th October 2011, 18:03

I just want to come up with some smart "default" behaviour - i will probably just quickly check how Haali behaves, and try to mirror that as a default so people get what they expect. :p

Sadly, that probably makes the most sense for defaults. sigh :)

nevcairiel

16th October 2011, 18:05

Ok, let me boil this down to something that hopefully makes sense. :D I only ever want "automatic subs" in the case of forced subs being found for my preferred language. I NEVER want to see default subs until I explicitly decide I want them. Does that help? :)

Thats of course possible, it does however require the advanced mode, because the "simple" mode will also activate streams marked with the MKV "default" flag.

You would do something like this in advanced mode: "*:eng|f", and repeat for all languages you understand. :p
(Quick explanation: *:eng|f -> For */any audio language, try to find a english subtitle track with a forced flag")

SamuriHL

16th October 2011, 18:07

Thats of course possible, it does however require the advanced mode, because the "simple" mode will also activate streams marked with the MKV "default" flag.

You would do something like this in advanced mode: "*:eng|f", and repeat for all languages you understand. :p

Perfect. I'm completely good with that. Thanks!

SamuriHL

16th October 2011, 18:09

Can you do negatives in your advanced mode? Something like this:

*:!eng|fd

Meaning if audio language is not english, show forced or default subs in the preferred sub language. Is that doable so I don't have to go through all non-english languages? :D

nevcairiel

16th October 2011, 18:34

Negations are not possible, but also not required, imho.

You could do something like this:

"eng:off *:eng|fd"

English audio: turn subs off, any other audio: english forced/default subs"

The expressions will be parsed from left to right, trying to find the first match.
I'll document the advanced mode when i'm done writing it. :p

SamuriHL

16th October 2011, 18:37

Sounds good. Subtitles have always been a pain. MC16, IMO, doesn't make it all that much easier with its overriding logic. I wish it would just let me use LAV filter's logic. I may request an option to do that. :D

masonwu

16th October 2011, 18:42

haha l like your works ,guy ~ i use it to make up my HTPC very well

Mercury_22

16th October 2011, 19:10

IMO the DEFAULT subtitles setting (behavior) should be as simple as:

1.If no languages are configured

a) Show only forced
b) If no forced exist DO NOT SHOW subtitles

2. If languages are configured

a) Show the matching for those languages & forced
b) If no matching for those languages DO NOT SHOW subtitles

And "that's it that's all" ! :)

But if you really, really, really :p want to satisfy everyone ( which I don't recommend simply cause you CAN'T :D ) :
3. Advance
a) "everything" user customizable :p

P.S. What I mean is more of "Do not show subtitle unless the user really wants them !"