Server: Picture serious degraded 2.2 vs 2.1

[MauriceVeraart]

yesterday we did a compare with 2.2 vs a other playback medium.
we noticed a "fuzziness" on casparcg

this morning I compared 2.1 vs 2.2 with a screen consumer.

what you see in the pictures is a clip brought back to its native resolution with the mixer command(ARAL)
and a native pixel on pixel testpatern as a PNG not touched by any effect
(click image to see real diffference)

all this tested with the screenconsumer in windowed borderless and bordered mode.
there is a significant difference between 2.1 & 2.2

LS

[ronag]

@dotarmin @Julusian related to intel GPU PR?

[dotarmin]

I don't know, I can compare between two v2.2 versions, one with Intel GPU PR and one without. I will do that in one hour when I'm back.

[MauriceVeraart]

tests done with
88171e16d (ARAL = custom build based on this)
and
abdbcc196 (pixel on pixel off test)

[dotarmin]

@MauriceVeraart can you also do same test as you did with latest v2.2 and version from 2 May available at builds.casparcg.com?

[MauriceVeraart]

Just did exact same results

[dotarmin]

Just did exact same results

Thanks.

I have done some quick tests my self and neither can't I see a difference between the Intel GPU PR and the builds before it. I need to run some more tests before I say more but I will do some more tests and upload images from the output.

/Armin

[jesperstarkar]

This must have something to do with deinterlacing?

[ronag]

@jesperstarkar possible, we always deinterlace now, I don't remember which issue it was discussed. Someone was quite sure that bwdif wouldn't do anything if the content wasn't interlaced.

Could try to remove bwdiff and run again to see if it is still the case.

[premultiply]

That was me.

[ronag]

I'm a bit unsure what to do here, I like the always deinterlace method and I can't think of any foolproof alternative. Only alternative is to trust the media meta data in regards to whether it is interlaced or not.

[premultiply]

It is not shure at all that this is caused by the deinterlacer here.

[MauriceVeraart]

Maybe make always deinterlace a option in config ?
like interpret content always as progressive
always like interlaced
or by meta data ??

in my context I am pretty sure I will always get progressive but that's of coarse different per user

[dotarmin]

@MauriceVeraart It's not verified it's caused by the deinterlacer.

[MauriceVeraart]

true but what we see it could be.
no clue how easy / hard it is to switch it off just to try

[premultiply]

What it interesting here is that it seems to be blurred in H and V direction.

[Julusian]

and a native pixel on pixel testpatern as a PNG not touched by any effect

If I understand correctly it is also affecting pngs, so I dont think the deinterlacing could be causing this.

@ronag Was the opengl texture sampling mode changed or something like that?

[dotarmin]

@Julusian If I understand correctly it is also affecting pngs

That's correct, or at least what I could se when looking at the difference (in photoshop) of two snapshots with v2.1 respectively v2.2.

[MauriceVeraart]

yep also on png

see attached grabs with 2.2 and 2.1 with maybe a more usefull testpattern

to be clear last one is 2.1
grabs now made with casparcg.
maybe this can clear out the screen consumer to be the victum

[ronag]

If I understand correctly it is also affecting pngs, so I dont think the deinterlacing could be causing this.

I'm not 100% sure pngs are not going through ffmpeg.

[MauriceVeraart]

almost same result by playing the last png with the HTML producer

don't know if this helps

[premultiply]

Then it is not caused by the deinterlacer.

[dotarmin]

I did some more testing today to confirm the results I got last week. I played AMB and when it stopped on the last frame I did a channel snapshot. I did this first with the latest v2.2 (89ff914) and then with v2.1 Beta 2. When I had the channel snapshots in my hand I did a diff on those in Photoshop.

v2.1.0 Beta 2

v2.2.0 89ff914

Difference

I don't know if this helps. Also, I used decklink consumer instead of screen consumer. Testing done in 1080i50.

[premultiply]

Freeze-Frame-Video is not best for comparison here as different deinterlacing and upsampling is used in 2.1 and 2.2. And I am not shure they freezed on the same frame/field. But for me this looks ok (as expected) in this case on the first look.
But again this is not the best test material for this issue.

[dotarmin]

@premultiply Freeze-Frame-Video is not best for comparison here as different deinterlacing and upsampling is used in 2.1 and 2.2

You're right, this maybe wasn't best test material but I did same tests last week but by using a png image instead with a solid background and a large white text on it. Result was the same. I can upload those images instead if it helps? ping @premultiply

Thanks!

[MauriceVeraart]

ok did the same as you dotarmin but with the testpattern i used earlier

next image with the levels a bit changed to see whats going on more clear

[MauriceVeraart]

looks like the image is feathered out 1/2 a pixel on each direction a bit like the effect of a low pass filter

[baltedewit]

@dotarmin these were at the same channel resolution right? At 300% zoomed in, the image from 2.1 looks a lot sharper:
2.2

2.1

Not sure if related to this though.

[MauriceVeraart]

You did the zoom externally ??

sorry I have te read first ;)

[dotarmin]

@baltedewit these were at the same channel resolution right?

That's correct, 1080i5000.

[baltedewit]

You did the zoom externally ??

Yup. And look at what happens when we boost the brightness on the diff Armin made:

_note: maybe some of the finer noise is caused by encoding on GitHub's side, but that doesn't take away there is still a difference_

That's correct, 1080i5000.

I thought so, just wanted to be sure it didn't occur from a resolution change

[dotarmin]

@baltedewit, as @premultiply wrote in an earlier post:

Freeze-Frame-Video is not best for comparison here as different deinterlacing and upsampling is used in 2.1 and 2.2. And I am not shure they freezed on the same frame/field.

But I get same results with a full-frame PNG-image with a solid background and a text. The result is similar to the text "CasparCG" from the AMB samples posted.

[MauriceVeraart]

Looking in to this a bit more
i tried if i could make the picture sharper by shifting it 1/2 1/4 or 1/8 of a pixel

to see if it is a phase problem of some sort
(yep that's a analog approach)

in 2.1 i can make the picture look like 2.2
but unfortunately i could not do the same the otherway around

could it be that the picture is shifted some where in the chain say 1/2 or 1/4 a pixel
(maybe due too colorsampling conversions)

[ronag]

If someone could go back in commits to where this problem starts to occur that would be very helpful.

[dotarmin]

@ronag If someone could go back in commits to where this problem starts to occur that would be very helpful.

I'm not at the office this week so if someone else could test the first v2.2 build to compare it would be great. Otherwise I can try to do this when I'm back.

[MauriceVeraart]

Will do when I find sometime

[MauriceVeraart]

Ok first test was with the first available version on builds.casparcg.com
casparcg-server-7ba5e475f07ab7b6d1391097b04059b470457095-windows.zip 28-Feb-2018 15:50 142M
Unfortunatly the result is the exact same

[Julusian]

@ronag This started with the commit https://github.com/CasparCG/server/commit/e018e25acd424cea1ec0b1f2fe784a4ba5850382
In particular this block:
https://github.com/CasparCG/server/blob/3c77a8378fde2384200a6494e2204a3acecd1b99/src/accelerator/ogl/image/image_shader.cpp#L265-L268
By changing that back to return texture2D(sampler, coords); there is no longer any quality loss, so that algorithm needs revisiting.

[premultiply]

Is it a problem to simply change (revert) it to texture2D() again in source to fix it? @ronag Why was it changed to textureBicubic()?

[ronag]

Why was it changed to textureBicubic()?

Higher quality. But there is no significance. Can easily be reverted.

[ronag]

Try with latest master.

[MauriceVeraart]

will do later this week
thanks

[dotarmin]

I would love to help testing this but I'm away for a couple weeks on vacation. Please let me know if I can do anything else to help out.

Best regards,
Armin

[ronag]

I'm closing this. Please re-open if the problem remains.

[dotarmin]

Works like a charm now!

[MauriceVeraart]

Indeed problem solved