http://virtualdub.everwicked.com/index.php?act=ST&f=6&t=6136

feel free to comment here,or at vdub forum,doom9 forum ( http://forum.doom9.org/showthread.php?threadid=71485 ) or at our emails....

I'm now reading it.

The link to the FFVFW.avi seems to be dead.

seems you're right...i'll inform him....

Thanks! It should work now ...

Interesting test. If I have the time, I use pixiedust myself.

I personally found, that pixiedust aids compressibility alot (compared to peachsmoother). It may be a bit more blurry, but if you are limited in bitrate, the overall final result will be better imo. I'd rather have a little more blurryness than a lot of visible macroblocks.

Not saying that your test contradicts this in any way, it's just my personal observation I'm sharing with you.

we liked pixie best too....we hated it's speed though....(but that doesn't make it less of a winner)

ie i don't have time nor the cpu for it( )

pixie will make files more compressable...sure....

Very interesting!

I wonder about one ranking though:

Only looking at the screenshots, I think that Overlay preserved more detail, e.g. hair. (unless you say: "Hey, look at the crisp edge of this block!" )

Do you know, how these filters would behave with h.263? Does the difference become smaller?

uhm...?
i don't think...
overlay came quite blurry as his source is very noisy....hair looks sharper with peach...noisier,but sharper too....
for me overlay was untolerably soft...

block edges?well.....i agreed with wilbert that there's no point in making the test without including the codecs..at first i wanted to do pure filters test (no codecs),but it doesn't make any sense....we don't have that effective lossless compressors...
however i think it is visible that overlay melted too many pixels....peach didn't as it tries to preserve edges....

i agree with wilbert's list...(but as you saw;
)
perhaps you prefer absence of noise but some loss of detail.....


h263 would just add another layer of spatial softening....

funny you ask that,i just did quantizer matrix comparatives.(pure spatial performance of codecs via avs' imagereader..)...perhaps i'll publish that too.....(my aim was to compare ffvfw's matrices to nandub by pure visual inspection)

i always seem to be short on web space....

good stuff!

looks like if we've got the time, multiple captures work best (i've tried this with VHS caps, but i only had 2 copies, recorded 12 years apart )

i've recently found a very light denoiser that gives an extreme compressibility gain while only minimally touching the picture...

search doom9 for "removedirt"

of course, i don't expect this to be added to the comparison, but it would definitely help for analog capturing in future.

oo...thanks..(at last... )


woohooo...yeah!

i should probably "remaster" my 14year old blade-runner vhs recording....
(i didn't told you;i have blade runner but it's not director's cut,and it's pan-scan...i didn't found the scene with "his mother" to be particularly interesting or particualrly hard for codecs.(it's a still scene with some motion when he kills interrogator)...actually i never had fog blocking on my encodings...and i surely didn't got it here....lil after this scene,there's deckard eating something in a public place....after that we see some car...in this scene (camera (or pan-scan?) going to the right) ffvfw did dark area better than xvid,and xvid swimmed a bit..(while deckard was eating and fog is around him,all is well..no problems for either codec).but that doesn't really matter,i'm just mentioning it without a good reason....i really don't wanna discuss xvid or ffvfw anymore......)


i was aware of it's existance,but didn't tried it...
hmmm.....i find his web awesome!all words and tables and no images!excellent!
nice light green background....
(i'm not kidding!)

i'll try this as i'm constantly in a search of mild denoisers that are fast...so far i used tom barry's "stmedian" when i needed real mild denoising(to preserve as much detail as possible..kept the temporal component real low....)
i'm also lookign for denoiser that can turn off(completely) temporal processing on a motion....(mipsmooth can,but it has lousy temporal component to start with...so it fells more like "always off"... )

thanks for the tip...


if i find it good enough to play in the first league,it will make it's way into the test....

tried it last night;i AM COMPLETELY BLOWN AWAY!
i used it for some 768x576->384x288(very poor vhs source) conversion...
this is some serious shit,let me tell you;and it should also be called RemoveNoise....

1-doesn't leave the "vertical lines shimmer" like peach does....peach trying to preserve edges is really a flop when it comes to vhs...removedirt is better

2-it will turn itself off on moving portions of the frame....if frame has still portian as well as moving one,it will denoise still portion and leave moving as-is.....(the very thing i wished for in my previous post)
(even if spatial processing of peach is very decreased,motion scenes will still look better with removedirt,as it's not doing any processing on critical places...peach is)

3-very fast;turning it off gained me cca. 1fps faster encoding only....

4-not mild at all;it's just as i love them;agressive where/when it's needed(still portions),mild where/when it's needed(motion)

few downsides;
1-if frame has vertical shivering(for example damaged vhs tape/poor tracking-you can replicate this error by moving tracking control on your vhs untill the image starts to jump up-down...),some weird effects are visible-most of the frame is still,while some portions "bounce"....i think this could be tweaked via motion-detection parameter tweaking...

2-here and there i noticed block misalignement errors,BUT i only noticed it when i compared filtered stuff to non-filtered side-by-side...otherwise i wouldn't noticed it,and errors were very small anyhow.....humans won't notice these at all on normal playback speeds....i wouldn't noticed it even on frame by frame basis if i hadn't the unfiltered clip in another vdub instance....

3-chroma denoise component is poorer than the one found in peachsmoother,but this can be compensated with cnr2 filter...

4-1frame after the scenechange(ie. the first frame of new scene) is not denoised,but...peach doesn't do well after scenechanges likewise....

these are a minor issues that don't bother me at all....
wonderfull.....

i used it alongside cnr2;
-resizer
-cnr2(mild settings)
-removedirt()

i wouldn't use any other temporal filtter alongside it(i'm so impressed and excited!)...
you said temporal smoother goes well with it,but i don't see the need for it....t.s. will only ghost on higher settings,and this filter is not spatial one so it can't smudge those ghosts produced by t.s....

regarding documentation;i think it needs scaling down;i'll try to do that if i have the time (i'll consider these two objections(vertical shivering issue and mismatched blocks) the most on my tries to tweak the settings)
also needs the packing of that msvcr70.dll(?)(although i didn't needed it,but seems some did)

big docs are good(for advanced users),but will only scare away others....

still need to do few more tests(i have some interference on VHF channels,that peach normally cleans pretty well..i'll see how does RD deal with them,and i'll do different resolutions(ie. higher ones); much attention will be devoted to detail preservation(especially on motion) etc.)

if all goes well(as it did so far) i'll suggest this filter as substitute/alternative for peach smoother.....

prior to this filter i was afraid to do temporal-smoothing on such noisy sources(if i wanted to preserve detail and avoid the ghosting) but this filter changed all that......



excellent tip;thanks again........


edit;i just tried 512x384 and it did well again;only motion artefacts are the ones produced by kerneldeinterlace....
our test clip (@648x480 1,2Mbit/s) is already encoded too....
i can't say that xvid will love this filter(as it leaves noisy stuff as-is on motion),but i will,that's for sure!
we'll see if wilbert would be willing to do QMF with RD as temporal component...

Could you post your avisynth script?

my script is too large to be usefull to most people(13kb file),but it's not so complicated;it's just that i put all things to one place,and switch on/off the things that are required/not required....(and also there are some comments in croatian.. )

important parts here were this;

notes;
for 768x576->384x288(or even 512x384) you may wanna try to skip deinterlacing alltogether...(the "blended resize" approach will produce smoother motion,although if you'll be inspecting video frame-by-frame,if won't look so smashing...i tested this on sports...search for differences on camera pans or so...)

cnr2 might even require still milder settings(residual chroma artefacts)...lower the 2 digit numbers in the cnr settings....

rd used on defaults,i said what i objected,so i would probbaly tweak motion treshold as the first thing...

bordercontrol to ease up the job to compressors on black borders(needed to perfectly preserve AR...my stuff has 0 AR offset every time...for example,to cover up vhs head switching...cropping it would result in non16 resolution and i don't want that......)

removedirt passed the VHF interference test,and also did well on higher resolutions(even without any parameter tweaking)
will test it's speed(direct comparison to peach) and after that i'll point wilbert to read this thread....
i think it's safe to mention this filter as alternative to peach(it will beat peach on vhs for sure...and on other(cleaner) sources it will match peach's performance...and all that with default settings....very nice...)

I will look at this filter The only problem is that I will be moving to a new appartment next week. So, it will take a while before I'm connected again (say two weeks).

i surely hope you'll get better cable there ( )

also shouldn't be much job to mention it in context of vhs,and as possible replacement for peach....(especially for film sources(?))

i said it passed the interference test,but i didn't said that it beat pech(as it didn't)
i think it's good for vhs as vhs has bigger grain noise(more like film's..for what this filter was made in the first place) and i still think that it'll beat peach on vhs....

peach showed strong protection against ghosting/excellent motion parts preservation on "spatial=0" and moderate denoising numbers(up to 50)
"moderate denosing numbers" doesn't mean that removedirt will beat it...infact i think peach will still yield nicer images in finer noise/interference than rd....will do some more tests to check this....

here are the times to encode the test clip("hand" sequence):

removedirt(mthreshold=50,pthreshold=20)--->3'00"

Cnr2(0,35,192,60,256,60,256)
removedirt(mthreshold=50,pthreshold=20)--->3'12"

peachsmoother(noiselevel=2.878,baseline=1.926,NoiseReduction = 50, Stability = 30, Spatial = 0)--->2'57"

somebody wondering about funny peach numbers?peach has a system to assess the noise prior to applying denoising;


no artefacts from either filter(i tweaked removedirt to avoid all the ghosting/block mismatches),lil bit better denoising performance from peach(this was tv-rip)
(less means faster encoding)

peach has one more thing to say for itself;it's easier to use;any frankly,rd doc is a mess for a rookies....fun for me,but mess for a newbies...

for 2 tv-rips tests i did(one with interference VHF cap,and one analog satv...both video(non-film) sources...) i would pick peach....for vhs i would use removedirt.....
for tv(film source) i will compare peach to rd some more...(rd should do well on film...)

seems like i just find miself a excellent vhs cleaning tool....

/ivo

Is there any denoiser that can be used at capture time? (Besides the built-in VDub noise filter. RGB conversion is not really tempting either...)

Today's machines are (nearly) fast enough for encoding MPEG4 1-pass in realtime, and I think it would be advantageous for MJPEG captures as well.

forget it.....



people,here's source-vs-peach-vs-removedirt comparison done on a tv-cap(film source)[still get that interference on VHF CH09... ] ;



http://i4004i4004.bizhosting.com/RD%20versus%20peach/

you'll be able to notice that peach left the moving parts alone(relative lo noisereduction threshold and no spatial processing)
you'll notice this even though i didn't posted clips...you'll see motion blur or context of the images....

many other rd settings were tried,but i wasn't able to get same results as with peach(ie same amount of denoising on still portions of the frame)
actually one of the issues i have with rd is that this filter has many "detection" related settitngs(motion detection and dirt detection) but has less settings that would influence HOW MUCH of denoising is done....


just look at this table

it's either motion or dirt detection...no setting with which one could steer denosing power directly...this is confirmed by tweaking paramters;
filter comes to a certain level of denoising performance and just stays there...no way i could change the amount of denoising on still portions of the frame (motion/dirt detection settings affect mainly what portons of the frame NOT to touch...)

i'm afraid this is NOT a way to improve upon peach smoother,and for sure i'm staying on peach for tv-caps....

for vhs,as i said...will use this filter(removedirt) as peach will leave to much of vhs vertical lines shimmer intacted...
"mthreshold" and "pthreshold" settings (of rd) need to be toned down in order to avoid ghosting/blocking(on motion...and i didn't mean that codec will block,but rather,this filter will produce blocking if it gets motion detection wrong)



/ivo

I tried out removedirt with the default settings, but quickly ran into some artifacts.

The movie I was experimenting with was "Black Hawk Down" special edition DVD released in Germany.

In one scene the helicopter blades turn blocky. Similar artifacts appear on almost every frame in the scene. Have a look (266kb PNG)

Just for kicks I tried pixiedust(limit=3) and the blades look fine.

yes,these are the artefacts...
as i said;


also,you shouldn't judge on frame-by-frame basis,but rather on normal playback speeds...(but i guess you did that..)

and as i said,peach will denoise more agressively,but won't have these blocks anyway...so i think for non-vhs sources peach might be a better choice actually...
(but then again,you're doing dvd and that has blocking noise etc.)

try both and see which one worked better...these are the 2 very best(and fastest.. ) temporal smoothers at the moment...

hehe. i'm glad this filter's getting some good use. must thank Kassandro for making it.

one thing which you've already noticed is that when conditions are wrong for it, it fails spectacularly (try any "rumble" scene in an anime).

i'm thinking of writing a script that chooses between unfiltered and filtered when this happens (shouldn't be too difficult... "overlay(filtered,unfiltered,mode="difference") > threshold ? unfiltered : filtered" in pseudo-avisynth)

i've been tinkering with de-blocking techniques (based on blindPP and some masking techniques) to get compressibility up some. not as good as removedirt, but when the 2 are combined

oh, and here's my removedirt line for dark foggy scenes where objects often disappear at the default settings (such as the all-too-familiar scene in blade runner)

removedirt(mthreshold=4*255,athreshold=18,soscillation=10,doscillation=255,dist=2,tolerance=12,show=
0)

it's pretty convoluted, and by now i've forgotten what the settings mean (yes, the web page is a little TOO much info, even for a data junkie such as myself, and i had to tune my command while the page was open)

i can post my de-blocker if you like, but it's a little early in it's life