Can some of you test Tom's SimpleResize filter?

[DDogg]

Could some of you with eagle eyes and HQ tvs test Tom's simpleresize filter for quality? I am looking at an encode going at 1.73 Realtime using his new filter (95 minute encode will take 48 mins). Quality looks great but I only have an old 32 in tv to judge from.
Edit: my baseline speed with standard bicubic on this encode is 1.32 RT

http://rilanparty.com/vbb/showthread.php?s=&threadid=14416

sample:
LoadPlugin("C:\PROGRA~1\DVD2SVCD\MPEG2Dec\mpeg2dec.dll")
LoadPlugin("simpleresize.dll")
mpeg2source("D:\waterboy\DVD2AV~1.D2V")
simpleResize(480,360)
AddBorders(0,60,0,60)

[int 21h]

I'll look into using it some, I wasn't sure of the quality differences, so I'll have myself and some associates test it out for quality purposes.

[DDogg]

great! I know you and your "associates" know your stuff :)

[Mozart]

hi,

there is a good increase of speed. Testing with "shrek", R4, NTSC:

bicubicResize: 1.35 RT
bilinearResize: 1.48 RT
SimpleResize: 1.72 RT

The result seems almost the same in my PC. I will test the quality in my TV tonight.

DELL precision304, P4 2GHz, 1 GB RAM.

[DDogg]

Looks like our machines are very similar in encoding speed. This could be useful to know in the future. I am looking forward to your quality eyes report :)

[dvd2svcd]

@Mozart off topic:

..never argue with idiots..they drag you down to their level and beat you with experience..

So true, oh so true.

[Mozart]

@Ddogg

oh no. This machine is not mine. I'm in the lab of my college... This machine is too expensive to me.:(

[trbarry]

Thanks guys for checking this out. I don't have golden eyes myself except maybe for a very small aquired talent for being bothered by certain deinterlacing artifacts.

Appreciate any comments.

- Tom

[gerti67]

I did some testing too (1.0.6 Build 2 Pre 1):

0.877 CCE RT with BicubicResize
0.642 CCE RT with BicubicResize in Safe Mode

0.964 CCE RT with BilinearResize
0.677 CCE RT with BilinearResize in Safe Mode

1.071 CCE RT with SimpleResize
0.741 CCE RT with SimpleResize in Safe Mode

So as a result you can see that on my system (1.2 GHz Duron, Win2k) the encoding speed increases by about 10% from BicubicResize to BilinearResize and by about 10% from BilinearResize to SimpleResize.

Another result is that by using the Safe Mode i get a drop in encoding speed by about 30% with each Resize method.

Now to the visual results, tested on PC and on a big TV:

I liked the results of the SimpleResize very much, because i think the visual result is something between bilinear and bicubic. The picture is not so faded like with the BilinearResize but the Ghost Shadows and the Mosquito-Effect is not as strong as with the BicubicResize, so i think for me it's a perfect compromise even if you look at the increased encoding speed.

I did testing this with only a few chapters of "Evolution" (PAL,Region 2) and the results may be different dependiing on the source but i think using the SimpleResize is a good step. I hope dvd2svcd will support this method in the next release so that many users can test it without having to edit the Avisynth Script, and we get more results so.

P.S. One thing i realized is that using the Safe Mode with CCE, the pictures look more colorful (a little bit to much for me). Can someone verify this effect?

HTH,
gerti67

[DDogg]

gerti67,

Thanks a ton for testing and the data. Very useful indeed. DVD2svcd has added the simple resize to the next build.

[24hourloop]

Originally posted by gerti67
I did some testing too (1.0.6 Build 2 Pre 1):

....
HTH,
gerti67

Did the same test, also Athlon 1.2GHz. Absolutely identical results. 10%. Nice picture quality (Cannot distinguish on my 35" from Bilinear).

[DDogg]

SimpleResize is certainly looking promising. Hopefully we will get some data from Mozart and int 21h when their time allows them to do so. That should give us a good cross section of opinion.

[bradnopit]

i tested the simpleresize filter and the results are unreal

test system
WiN XP PrOFF vers 5.01.2600
EPOX 8kha+ SE Raid
2 ibm 60GXP raid 0 stripe 8k
AMD 1900+ 1 gig pc2400 memory at critical 2 2 2 , 1T cmd

speed 1.920 (without simpleresize,1.65 bicubic)

AMD 1900+(@145 fsb)looking at an speed of 2.019 now

have to check the quality though encoding gladiator as test max 2600 avg 2400 min 2000 4 pass VBR

[markrb]

bradnopit with those settings you are almost better off going CBR of 2500 or so and the encode would be about 2-3 times faster.

Mark

[Clixo]

just tested it wit ts and 3 times vbr:
too much mosquito noise around objects. with bilinear resize just fine no noise around. ok the speed increases a bit but i rather spend 2 or 3hours more making and encode that i will look in the future several times and everytime that i see it i dont find any noise than to make it faster and with poor quality. with this i dont mean that it dont worth to trie, as ddog said it looks promissing... i will wait for another release.

[trbarry]

with this i dont mean that it dont worth to trie, as ddog said it looks promissing... i will wait for another release.

I likely won't be adding any filtering to it as that would mostly just turn it into BilinearResize and we already have one of those.

I wrote this with the idea that:

* I was almost always downsizing (it probably won't upsize well)

* Much of my material (HDTV caps, DVD) was very clean and already somewhat filtered.

* Bilinear resize looked somewhat softer than it would have to be.

* I could make it go faster if it was limited to a fixed 2 tap filter.

But thanks for testing it. I'm adding some more SSE2 and some SSEMMX optimizations now.

- Tom

[gerti67]

@clixomano:

Did you test it on a TV or just on your PC?

@all:

When i was testing it, the BicubicResize and the SimpleResize in fact produce Mosquito-Noise. But from my point of view this only happens around medium objects moving with medium speed and with the material i used for testing (different chapters covering almost all type of scenes) this rarely happend.

There is no noise around big objects in the foreground and the most important parts of the scene because movie scenes are almost perfectly planned by the regisseur and so there is always little movement with small or medium objects. When there are scenes where small or medium objects are moving then the total impression of the scene is more relevant to the watcher so you try to get the whole scene and not to just try to make out some noise around that little person there or this car there. (So from my experience you only notice this noise when you are looking for it). With fast moving objects you aren't able to see this noise because it all happens to quick for the watchers eye so you have to watch it in slow-motion to make it out and i don't do when watching movies.

It is true that there is almost no noise when using the BilinearResize but for me it softens the whole picture way too mutch. Objects in the background and even the whole background scenery are looking some kind of blurred. With big objects in the foreground it is almost ok but not comparable with the bicubic or the simple method of resizing, they look much better then.

For me, using the BilinearResize reminds me of watching a video tape and that effect lasts for the whole movie so that i prefer using the bicubic or now the simple method because 50% of the noise you can make out on your PC is not noticeable on the TV. And the rest of the noise is in most cases only noticeable when you are especially looking for it.

For testing i made a SVCD with 4 different chapters and used all the 3 resize methods for each so that i got 12 MPEGs and burned them with Nero using a menu. So i could watch every scene with the bilinear, the bicubic and the simple resize method.

The results for me are: I didn't like the bilinear because of the reasons i described above and it was very easy to detect the bilinear scenes when using random play. To distinguish between the bicubic and the simple was not so easy but when watching precisely it was possible. The bicubic scenes just looked a little sharper than the simple ones. And with the SimpleResize there was a little more moiré with fine structure and patterns in the background - but you had to look close to make it out.

So at the end, this is not meant to offend you or anyone who prefers the BilinearResize, i just tried to describe the effect i was able to see because DDogg and trbarry asked for it and i had a little time for testing.

I think for the ones who preferred the bicubic method the SimpleResize is a very good alternative because it gives you an increase of encoding speed over the BicubicResize of about 20% with almost the same visual result. The ones who preferred the BilinearResize will not be so happy with the simple method of trbarry.

But for me the SimpleResize is a good one, so a lot of thanks to trbarry for making it and DDogg for posting it here not to forget dvd2svcd himself for this great program.

HTH,
gerti67

[DDogg]

gerti67,

Again, outstanding job. It shows you really went to a LOT of effort to do useful and objective testing for the group. It is really appreciated by me, and I am sure, many others.

DD

[Clixo]

@ gert67
only on pc but i will check it on the tv ... i see you poin and i will check it once again.

@ trbarry

i will test the new version with :SSE2 and some SSEMMX optimizations as soon and as you release them

[24hourloop]

Originally posted by gerti67

It is true that there is almost no noise when using the BilinearResize but for me it softens the whole picture way too mutch. Objects in the background and even the whole background scenery are looking some kind of blurred. With big objects in the foreground it is almost ok but not comparable with the bicubic or the simple method of resizing, they look much better then.


Gertie,

when you examine the test shots whose links are posted elsewhere here (I am sorry, I can't remember which thread) you will see that while BicubicResize seems to render a sharp picture but it is merely faking it. The sharpness it produces is simply not there in the original.

Another big drawback of of BicubicResize is the fact that due to the introduction of this 'sharpness' you gobble up overall compression bandwidth. The encoder has to work that much harder to compress the picture than if you have a 'soft' (but true to the original) image.

It would appear counterproductive to me to use a temporal smoother on the one hand to squeeze out more out of the available bandwidth and then to burn it again by sharpening.

Just my $0.02.

[gerti67]

@24hourloop:

I remeber that thread(s) about tho whole BicubicResize problem. From my point of view it is true that the BicubicResize (and the SimpleResize) are creating an artificial sharpness but i think you have to consider that you loose 1/3 of the original picture information by resizing it. And while watching it on TV from my technical understanding the DVD-player has to scale back (stretch) the SVCD source (480x576 for PAL) to the TV's resolution (720x576 for PAL) so that you don't get those "egg-heads" (correct me if i'm wrong).

So here is the point, by stretching back a bilinear resized picture the softening (blurring) effect is becoming stronger than it already is. So to use your example: BilinearResize will make the image softer than it is in the original. (As i already said it reminds me of seeing video-tapes.)

From my experience using the BicubicResize will create an artificial sharpness but by watching it on TV (with the stretching from the DVD-player) it is closer to the original than with the BilinearResize. In fact that is what i get as a result on my TV and DVD-player combination (that can differ from other combinations for sure).

And to come to the compression thing, i encoded a lot of different chapters (mostly short ones about 1-3 minutes each) and used always 4-pass VBR with the three different methods of resizing and i can tell you that the results were sometimes confusing as for some chapters the files with the BicubicResize were smaller than the ones from the BilinearResize, so that you can't generally say that this or that method is better for compression because it depends on the scenes. Not to mention that the sizes differed only by about 0.5-1.0% with these short clips. I know that can be different when encoding a whole movie but i think by using 4-pass VBR you can really forget about that compression question.

I also tested different combinations for the b and c value of the BicubicResize some time ago and found out, that the combination of 0.2(0.15)/0.6 is the one i prefer over all other settings (very little Mosquito-Noise and a little softer than with 0/0.75). By using these values i got nearly the same file sizes as with the bilinear method but it looked better on PC and on TV.

But it think is almost impossible to argue about taste because beauty lies in the eye of the beholder and the whole encoding thing with its tons of possibilities and parameters to fine tune is not making it better for anyone so you have to find it out by yourself. And i think the experiences from others will make it less difficult for the ones that are interested in this stuff.

Greetings,
gerti67

P.S. I never used Temporal Smoother, so you must have gotten that wrong.

[24hourloop]

I didn't mean to insinuate you had used the temporal smoother. Somewhere here is a lively discussion on that subject matter, suggesting that compression improves with a temporal smoothing filter.

While I am not an expert at this, this is what Ben Rudiak-Gould (the guy that wrote/writes Avisynth) has to say about this:

If you are magnifying your video, you will get much better-looking results with BicubicResize than with BilinearResize. However, if you are shrinking it, you are probably just as well off, or even better off, with BilinearResize. Although VirtualDub's bicubic filter does produce better-looking images than its bilinear filter, this is mainly because the bicubic filter sharpens the image, not because it samples it better. Sharp images are nice to look at--until you try to compress them, at which point they turn nasty on you very quickly. The BicubicResize default doesn't sharpen nearly as much as VirtualDub's bicubic, but it still sharpens more than the bilinear. If you plan to encode your video at a low bitrate, I wouldn't be at all surprised if BilinearResize yields better quality.

I am not sure I can add to that.

I guess in the end much of this is splitting hairs anyway. The ultimate test is whether you like the picture. If you like it you got it right, if you don't you made a mistake. Period.

[Mozart]

well... first, sorry by the delay of result.

My results are exactly the same of gerty67 and, after such detailed post, there is no word to add. All that I can say is: simpleresize is the best option, because of its net result when you are looking for a good speed(x)quality parameter.

btw: there is no reason to use TS if you are doing 3 or more pass. Only use it if you are doing 1 pass vbr or CBR. If someone want to use an average bitrate smaller than 1300, the option is to use CCE_patcher and change the quantizer matrix. Some matrices are very good, allowing you to fit a whole movie into 1CD of 800 MB, with a quality better than VCD. For example, "shrek" seems very good in my Sony Trinitron (at least for me;))

[int 21h]

My preliminary tests indicate that this is pretty good quality, and the speed improvement is somewhere between .2 to .5 depending on initial encoding speed. After sending around the crew, we think its a good improvement over the original two options, and I think we'll probably use it in the future.

Also looking forward to more SSE2 enhancements to make this even faster :)

[trbarry]

Thanks everybody for all the work you've done on this. I've added a bit more SSE2 and some SSEMMX (Athlon, P3, 550mhz+ Celeron) optimizations. I don't have any timing results yet but there should be at least a modest incremental speed improvement for any box > a P2 or K6-III. But probably nothing spectactular.

I hopefully will have it out by this evening or tommorrow at the latest.

- Tom

[trbarry]

Yesterday I promised a new SimpleResize with added P3/Athlon SSEMMX support. The good news is I got it working last night. The bad news is it doesn't seem to give any noticable performance improvement over v 0.1. Since the only other improvement is a better WarpedResize formula I think I'm just going to hold on to it and tinker for awhile.

I don't yet really understand what's going on (but don't like it).

- Tom

[gerti67]

Hi Tom,

don't worry too much about it, as the version you already posted is a really good speed improvement over the BicubicResize.

Right at the moment i'm just testing a few parameters with CCE (Image Quality Priority, Anti Noise Filter) to get a visual result that's right between the BilinearResize and the BicubicResize (0.00/0.75) or to say better that's close to the BicubicResize (0.15/0.60) that i prefer.

So just keep cool, sometimes it's better to leave the coding for some days and after getting back to it, it sometimes hits you right between the eyes and you find out what went wrong with the new version. (Got that from my own experience as i'm into database-programming with delphi which is surely not so tricky as your project with this SSEMMX optimizations, in fact i think that is really the hardest stuff you can go for when programming.)

So many thanks for the work you've done with the SimpleResize it is really appreciated by me.

Greetings,
gerti67

[chainsaw135]

I just used simpleResize last night, WOW its awsome, it has speed and looks so good i was amazed..keep up the good work.

[Clixo]

just to say that i remove what i said in my last report just used simplerisize with donalds graph telecide ( deinterlace ) and the result was perfect. so justo to say TOM good work !!!

[DDogg]

"Somewhere here is a lively discussion on that subject matter, suggesting that compression improves with a temporal smoothing filter. "


24hourloop, are you suggesting that temporal smoother does not increase compressibility? If so, please educate us on how you made your conclusions. I would be very interested in them.

[trbarry]

Well thanks to DmitryR for some useful tuning suggestions I managed to improve performance a little bit. Maybe it's approaching a point of dimineshing returns here. But I did get maybe an additional 3-5% speed increase for SSEMMX P3/Athlon users, based upon the total time of a job with no other filters. And maybe a tiny improvement for P4 boxes.

Version 0.2 of my SimpleResize Avisynth filter is now out, see: www.trbarry.com/SimpleResize.zip or www.trbarry.com/Readme_SimpleResize.txt

I posted more about this in the Avisynth forum in this thread (http://rilanparty.com/vbb/showthread.php?s=&threadid=14416).

Thanks again to all here for the testing and feedback.

- Tom

[gerti67]

Hi Tom,

just tested the new SimpleResize.dll with my default testing chapters and i'm afraid to say that for my 1.2 GHz Duron there is nearly now speed increase. (0.002-0.005 in CCE) Although it has the new Morgan-Core that supports SSEMMX like the new Athlon but perhaps my memory bandwith is the bottleneck with my system because it's a KT133 chipset with PC133 CL2 SDRAM only. So this can be different for others using DDRRAM with a KT266A perhaps.

Anyway, thank you for working on it i now use the SimpleResize as default setting.

Greetings,
gerti67

[trbarry]

I maybe spoke too soon about Athlon speed improvements as this has also been reported on the Avisynth forum. I did the timing on a P3 and P4 but assumed the Athlon would pick up the P3 improvements because it supports those SSEMMX instructions.

There are two possibilities here that I can think of offhand.

The first is that while the Athlon supports SSEMMX instructions it doesn't really make them go any faster.

The other is that Avisynth does not report the SSEMMX support on Athlon processors (or I'm misreading it) and so it is not even executing the new code. In that case I can fix it.

I'll know soon. At least I didn't make it go any slower or catch fire or anything. I guess it's nice it runs faster on my P3 but I don't usually even use that machine for encoding anymore. :confused:

- Tom

[trbarry]

No other particular performance improvements but, for those of you working with interlaced source, I just released SimpleResize v 0.3 which includes a new InterlacedResize function.

I discussed it more in the Avisynth SimpleResize thread (http://rilanparty.com/vbb/showthread.php?s=&postid=78319#post78319)

And I hope this post is okay. Would it violate the cross posting rules? There was already a discussion of SimpleResize here and you guys have done most of the testing, but I'm not sure.

- Tom

[dvd2svcd]

You are not crossposting, and I for one am glad you do post here too. It seems to me too that this is where the most testing have been done. Keep up the good work Tom, it's a great plugin :)

[SiliconSoul]

dvd2svcd

i have a question for you!

the bitrates page is for which resize? because for the movie im doing it said i could use a max avg bitrate at about 2600-1900 do i did it at 2500 with max 2600 and audio 192 and it plays fine with the simple resize... the problem is that for the 2600-1900 it said it would take 2 cds but it took almost 3 full cds with a 10 pass vbr with cce. so if the simple resize does not compress as well then you need to update the bitrate page for bicubic + bilinear then bitrate for simple resize!

btw the quality and speed is GREAT just the size i got was unexpected!

:-)

any help, idea, suggestions, etc would be great!

sorry about used the old thread just did not want to start a new one!

[Mozart]

of course you should start a new one, because this question is almost off topic. ;)

the resizing method has nothing to do with the average bitrate used by CCE. If CCE uses an average bitrate equal to 1900, the movie should be smaller than 107min. It seems that you are messing with those average values... Could you read the Q&A thread? There you can find some informations about how does DVD2SVCD works with the bitrate tab (maybe Q38... I don't remember)

[mrbass]

I've done about 10 or so dvds using simple resize. I think it's pretty good and acceptable. I do notice mosquito noise around people (not close-up though) when they far away. It's really hard to see unless your looking for it. I'm going to stick to BiCubic but if I needed my computer and it didn't finish while I was sleeping I would definitely use Simple Resize and not feel bad about using it.

I wonder which is better though. a) VBR 3 pass Bicubic or b) VBR 4 pass simple resize (avg bitrate of 2000 for both)? I think method a would be quicker but I'm not sure if it'd better quality or not.

[DDogg]

I wonder which is better though. a) VBR 3 pass Bicubic or b) VBR 4 pass simple resize (avg bitrate of 2000 for both)? I think method a would be quicker but I'm not sure if it'd better quality or not.

Mr. B, excellent question!
We await your research and answer :):)

Really is an interesting Q.

[SiliconSoul]

do you think certain dvd players like a different resize better?

bicubic ... bilinear... simple resize?

and you think a 3 pass bicubic is enough?

i have the machine to do a 5 pass overnight so it does not matter just wondering although.

[SiliconSoul]

bump?! :-)

[DDogg]

I don't think the player knows or cares and reproduces what it is given to reproduce.

Bicubic 0.0,0.50 3 or 4 pass is the way I always personally go (assuming bitrate is up above 1700-1800). If the bitrate is 1600-1800 I would do a four pass. If it is above that I would do a three pass for a general guideline. If the movie was very bright with saturated color and action it never hurts to give it an extra pass if you have the time.

[SiliconSoul]

this should be put in sticky... and all the good info from this thread!