.. excellent work on file size reduction with temp smoother .. [Archive]

Doom9

2nd January 2002, 14:21

well... I can't quite agree with "smoothing doesn't reduce crispness".. every time I used any smoother filter I know of I could tell the difference... and divx4 already smoothes the picture a bit too much for my taste without any additional filters. To be perfectly honest.. from a crispness point of view divx3 is superior to divx4. If I get blocks.. I don't use a smoother, I reduce the resolution because the former is much more visible.

Acaila

2nd January 2002, 14:53

Temporal Smoother works between frames unlike normal smoothers/cleaner which work inside frames and as such has a much less visible effect on crispness. Using it before resizing results in an even further decrease on visible effect.
I think it's a great way to reduce total filesize with as little effect on crispness as possible, but that's what everyone has to decide for him/herself.

Peters

2nd January 2002, 15:52

I agree with Doom9 , i'm waiting for Xvid development, hoping a better post-filter than Divx4 and a real coherent 2 pass, hello Blacksun! :)

MaTTeR

2nd January 2002, 20:15

ChristianHJW,

Thanks so much for posting this link. I gotta try this :-)

@Acaila-
Great work and testing. Thx alot!

ChristianHJW

3rd January 2002, 12:51

... the key here is to use the filter before resizing .... TheWEV made me a very happy man by adding this feature as a standard into GKNot .

Yes, there is a loss of crispness, but when resizing to normal resolutions for 1 CD rips ( and this is what its ment for, i never use filtering for 2 CD rips ) like 560 * 240 ... 608 * 256 the effect is only marginal, even smaller than using a different resizing filter, but will give better file size reduction overall.

serbersan

3rd January 2002, 18:27

...and I'm not sure the effect of smoother filter was so bad for 2 CD rips.

I've done a 2CD rip of Proof Of Life at 640x256 tempsmoother(1) and I'm very pleased, the overall quality is really great I doubt it could be so great without the filter. But it's only an opinion.

I'm doing 2 encodings of 2CD from "Dr. T." I want to see how the subjective is the opinion of doom9 about using lower resolution instead use a filter.

First 2 CD rip is 640x256 TempSmoother(1)
Second 2 CD rip is 608x240
Duration 1:55 min aprox || Audio 120Mbytes, 145kpbs VBR mp3 audio.

In the first rip without tempSmoother(1) compresibility was about 65% I think this value isn't enough for me. Obviously if it was something like 75% I wouldn't use any filter.

Well tonight I finish the 10hours First rip and tomorrow the second rip. If someone is interested I could attach images of both encondings, when they finish.

Acaila

3rd January 2002, 19:19

If someone is interested I could attach images of both encondings, when they finish.

I am

b0b0b0b

3rd January 2002, 19:29

I can see how smoothing might help with a noisy source.

But if you have a clean source, aren't you just creating ghosting??

I don't buy it. Has anyone compared the compressibility of something with and without temporal smoothing? Assuming noise is out of the picture, isn't it unclear what the relationship between the amount of smoothing and the compressibility is?

I think an interesting experiment related to temporal smoothing would be to interpolate the motion vectors between resulting mpeg-4 frames to generate in-between frames. You could use bilinear, trilinear, or any interpolation algorithm really. The question is whether you end up with more perceived detail and better pans. Of course this is useful only for folks with high refresh rate display devices.

Edit:

I just realized that in many dvd's you can see evidence of film grain. I can see how tempsmoothing (w or w/o resizing) definitely would help with that.

Acaila

3rd January 2002, 21:52

If you have a noisy sourse then temporal smoother would have less of an effect than normal in-picture smoothing.
The trick with TempSmoother is that it reduces the difference between one frame and the adjacent ones, and since compression relies on storing differences only, it reduces file size. Using low enough settings will make it easier on the codec, but will keep visible distortions minimal.

Btw, my test was on a DVD source, so it was clean to begin with and still gave nice size reductions.

Peters

4th January 2002, 00:20

Originally posted by Acaila
If you have a noisy sourse then temporal smoother would have less of an effect than normal in-picture smoothing.
The trick with TempSmoother is that it reduces the difference between one frame and the adjacent ones, and since compression relies on storing differences only, it reduces file size. Using low enough settings will make it easier on the codec, but will keep visible distortions minimal.

Btw, my test was on a DVD source, so it was clean to begin with and still gave nice size reductions.

I read your post on the link given by ChristianHJW.
Interesting but i'm not sure it's possible to interpolate from a test with a bitrate of 6000 ...
In fact i'm sure that the percentage with a lower bitrate (for one CD rip) will be different..

serbersan

4th January 2002, 01:37

I'm totally agree with acaila, in some tests I did when I encoded movies the increase in compresibility with the filter depends of the movie.

With the little noise filter in Gordian Knot (it has less effect than tempsmoother(1) I think is like tempsmoother(0.5)) the compresibility increase is between 3-6%.

DJ Bobo

4th January 2002, 01:57

What about BilinearResize + TemporalSmoother(2,1) ?

I tested it on noisy DVDs, it make miracles, the picture looks so calm, natural and beautiful (I would call it, a charming picture :D). The loss in sharpness is unnoticeable, unless one compares directly the DVD-Picture and the DivX4-picture.

@ Christian
Why do you propose to use temporalsmoother before resizing?! I can't see the advantage of that technique.

Acaila

4th January 2002, 10:09

I read your post on the link given by ChristianHJW.
Interesting but i'm not sure it's possible to interpolate from a test with a bitrate of 6000 ...
In fact i'm sure that the percentage with a lower bitrate (for one CD rip) will be different..

Ofcourse the percentage will be different. The amount of size reduction is highly dependent on the type of movie scene and general noise too. With a lower bitrate you'll get more noise, and compression will increase. These percentages were only given as a guideline, not absolutes, to show that compression increased with higher settings and adequate compression increase can be achieved with minimal visible effect on crispness.

serbersan

4th January 2002, 16:59

Well I have finished my temporalsmoohter(1) rip, now I'm waiting for the 608xXXX and 576xXXX rip. BTW I have redone the compresibility check at 7% and it's 60% with 640x... without filter and with temporalsmoother(2,1) it's 69% (I'm surprised because the first minutes of the movie are very noise, in compresibility check it arrives to 500K) and the correct duration is 1:52

I've done some proves with constant quality 100% with temporalSmoother(1) and lowering the resolution and...

Being objective...yes, the image is more sharpness with 576x240 than with 640x256 with TempSmoother(1). But it hasn't better overall quality.

The image loss sharpness at noisy parts, like human skin (faces) and specially when those noisy parts are moving, faces of human skin moving, again. But the overall quality isn't better lowering resolution. And the differences are very subtle I'm very near of my 17" monitor.

However images without persons are really good.

More information on aprox. 1000 frames(I have more results but are about the same):

CQ 100% 15Mb 640x256
CQ 100% 12.6Mb 576x240
CQ 100% 11.7Mb 640x256 TS(1)

I'm waiting the encoder finish the job, but I think in those cases there isn't enough space and lowering the resolution to 576 isn't enough. I've seen some scenes of the rip ts(1) and there is a smooth effect in a wall but it's like the encoder hasn't enough bits to do it well and not like a temp smoother effect.

IMHO I think below 70% at 640x256 2 CD rips you could use tempsmoother(2,1) and the overall quality should be better(not faces but overall quality and faces weren't so bad). Or at least it's not as easy as lowering resolution.

Well like I said I'm waiting the results of the 2º and 3º rip before obtain a conclusion but, I think lowering resolution is good when 2CD were good enough not like this case when Ts(1) has less size than 576.

What are your opinions?

MaTTeR

4th January 2002, 20:40

Well I just ripped 3 different movies with my temporal setting @ 3. All were encoded for 1CD with audio averaging 128k and resolutions @ 560. All I can say is wow! Very big difference in quality indeed. The overall PQ (picture quality) was much better than the previous rips when I was just using the generic "Little Noise" selection in Gknot.

I really didn't notice any loss of definition using the filter but maybe it's just my eyes. Thanks so much guys for this great tip.

BTW- What about using a temporal setting of 2 when I encode amovie for 2 CD's? Would this have any effect?

b0b0b0b

4th January 2002, 21:19

Do you guys watch these movies on a TV or on a computer screen. If on a computer screen, do you blow it up to at least 1280x1024 on a 21"?

Just trying to get a feel for how you are gauging PQ.

MaTTeR

4th January 2002, 21:52

Originally posted by b0b0b0b
Do you guys watch these movies on a TV or on a computer screen. If on a computer screen, do you blow it up to at least 1280x1024 on a 21"?

I only watch mine on a 27" Tube TV @ 800x600 resolution. This obviously isn't optimal of course but I don't have the moolah for HDTV. This is how I judge my PQ. Watching it on a monitor is much less forgiving due to dot pitch & resolution.

serbersan

5th January 2002, 04:35

BTW- What about using a temporal setting of 2 when I encode amovie for 2 CD's? Would this have any effect?

Well I think a TempSmoother(1) is ALWAYS useful for 1 CD rips and at level 2 I think the same but I never go with level 4 and I have doubts about level 3 but maybe....

But well for 2 CD rips if you I think is better to lower the resolution, but the difference between the filter and lowering the resolution isn't so obvious.

As Matter says the overall quality with filter it's better but faces and other details have much more sharpness without filter and lowering the resolution. And I have used a TempSmoother(1). I think we could doubt between lowering resolution or TempSmoother(2,1) maybe (1) but I will never use level 2 for a 2 CD rip.

Well here I attach my results.
dres -> resolution 576x240 without filters
dfil -> resolution 640x256 TempSmoother(1)

serbersan

5th January 2002, 04:43

How I attach a file????

The Attach file doesn't work correcty, it allows me to attach a file but it doesn't show it.

MaTTeR

5th January 2002, 04:46

@serbersan

So you would always use at least a 1 with 1CD rips but why would make you use 2 for a 1 CD rip? If the film was more noisy you would use 2? What are you viewing your rips on to judge the PQ?

I'm currently encoding a 141 minute movie going for 2 CDs using a temporal value of 2 @ 608 resolution. I'll let you know how it turns out. Maybe I'll encode it a 2nd time using a value of 1 just to see the difference.

serbersan

5th January 2002, 05:02

Well MaTTeR, if I could attach the file I want you could view the difference between 576 without filtering and 640 with filter (1). And Yes the 576 has more sharpness in some cases but in others the overall better quality is for 640 (1) Well I choose 576 because the sharpness it's in zones of the image that are important.

For 1 CD I use TempSmoother(1) but if the movie it's easy to compress if you have any doubt use TempSmoother(2) but with level 3 in some action scenes or in faces the effect is very noticeable but if the movie is too large or difficult the overall quality will be good.

My advice is use TempSmoother(1)(dark movies) or (2)(not easy or very easy movies) but be warned with level 3 and I not recommend to use level 4.

And for your 141 minutes rip I think TempSmoother(1) will be good at least with 608, I think I used it with so long movies, or TempSmoother (2,1). In any case with 141 minutes I don't think TempSmoother was a bad choice only you have to see if the movie has much noise or is difficult or simply for other times....do a test with 1 and if the result wasn't good enough use more filtering.

I think for 2 CD rips first lower the resolution and then use filtering. But again, the difference will be very subtle unless the movie was easy to compress.

serbersan

5th January 2002, 16:58

Well I have to remade my post, finally I burned the filter movie because the overall quality was better, there isn't backgrounds with strange efects of moving(because the too much noise and the lower resolution)

Well I think lower to 608 is right but no less without do before some filtering.

I didn't do 608 without filtering because in 576 there was zones where appears smooth made by postfiltering because it hadn't enough bit to encode.

Sorry for you time.

ChristianHJW

5th January 2002, 22:50

@MaTTer : using temp smoother has nothing to do with removing 'noise' . Its a very effective method to improve compressability of the movie by removing delta information between consecutive frames ....

MaTTeR

5th January 2002, 22:59

@ChristianHJW

Hmm...well it certainly smooths out the graininess of the last 3 movies I did. I suppose that effect is just a byproduct?

serbersan

5th January 2002, 23:03

Are you sure ChristianHJW???

If I remember in the rip I've done of Dr T with TempSmoother, there was less noise in background and in some zones of the image, in some cases less detail in human skin. And the first minutes are very noisy.

IMHO the filtering in this rip is better than lowering resolution the overall quality is better(less mosquito noise for example).

Good Luck with your projects ChristianHJW

Peters

5th January 2002, 23:08

Originally posted by ChristianHJW
@MaTTer : using temp smoother has nothing to do with removing 'noise' . Its a very effective method to improve compressability of the movie by removing delta information between consecutive frames ....

hmmm, I suppose that putting temporal smoother as a noise filter in Gordian Knot is a mistake?!
:)

Acaila

5th January 2002, 23:37

@Matter

The reason why it removes "grainyness" is based on the fact that a certain pixel in one frame has the grain-effect, but the two pixels in the two adjactent frames are enormal. Hence it's being corrected and the "grainyness" disappears. BUT this doesn't blur the picture unless you use a very high setting.

@Peters

That's actually only half of the story.
If you push the edit button on GKnot you'll see that the noise filter is actually a combination of various decrees of TemporalSmoother and SpatialSoftenMMX. TemporalSmoother reduces the delta information between frames, SpatialSoften reduces the noise by blurring the image inside one frame.
Btw, used in conjuction they don't increase compression over using just TemporalSmoother alone. But SpatialSoften blurs a whole lot more than TempSmoother so don't use it unless you've got very bad noise.

Peters

5th January 2002, 23:50

@Acaila

SpatialSoften is not used by Gordian Knot if you choose little noise.
So don't say it's not a noise reducer!

TemporalSmoother(clip[,strength[,radius]])

This is a port of the TemporalSmoother filter from VirtualDub. It works in RGB32 & YUY2 pixel formats.
From VirtualDub help file:

This filter is an adaptive noise reducer, working along the time axis; it is most effective when the image is not moving much. Increase the filter strength to increase noise reduction, and decrease it to reduce speckling and ghosting artifacts. It is recommended that you combine this filter with a spatial (area-based) noise reducer for greatest effect

MaTTeR

6th January 2002, 18:01

My 141 minute movie turned out nicely using a temporal value of 2 with DivX 3(2 CDs). I'm going to encode it again with a value of 1 though.

I also decided to encode the movie with a value of 1 using DivX4. The results were unimpressive. I don't use DivX4 very often so I thought I would give it a try with Gknot and use the same Gknot settings basically as my DivX3 settings. Of course using DivX4, Gknot tells me I'll get 1227 bitrate compared to 1197 for DivX3 so I expected it would look alittle better. Personally, my eyes like the DivX3 version that I encoded with a temporal value of 2. However, now I want to encode it again with a value of 1 using Divx3.

I had my girlfriend view both DivX4 and DivX3 encodes without her know which was which. She picked the DivX3 version also. Overall, it seems Like the DivX 4.12 codec is much darker. It's almost like it tries darken an object in the picture that it feels is unimportant just to get better compressibility. That method would be fine but it's a little to agressive. I guess I now see why most of the pros/gurus here use DivX 3 instead of DivX4.12:)

Again, the temporal smoother setting just simply rocks! It gives a better overall picture without really losing any definition of the fine areas.

EDIT- Anyone notice that file size predictability isn't as good while using the temporal smoother? I'm having to add about 20MB to the total size selection in Gknot to fill the CD's

ThePanda

6th January 2002, 19:48

is it still believed that smart smoother should also be applied? (I apply temp. smoother before resizing (.6 bicubic) and smart smoother after resizing for my anime rips). also, what is this "2D Cleaner (Optimized)" I now see on Donald Graft's filter page? Is this comparable to or better than smart smoother?

Acaila

6th January 2002, 21:01

I've done some testing of TempSmooth with SmartSmooth in the past, especially on anime. What I saw was absolutely no extra compression of using TempSmooth + SmartSmooth over TempSmooth alone. However using SmartSmooth put a serious blur over my movies, even at low filter settings. Obviously I don't like using it anymore, I just stick to TempSmooth.

1 CD Movie -> TempSmooth 2 or 3
2 CD Movie -> TempSmooth 1

EDIT- Anyone notice that file size predictability isn't as good while using the temporal smoother? I'm having to add about 20MB to the total size selection in Gknot to fill the CD's.

File size prediction doesn't take into account the effect of TempSmooth, because it's different for every movie (and probably because it wasn't tested thouroughly before GKnot was put on the market). And since file size predictability is already off for DivX4 and few extra MB won't make much of a difference :D