did you compare? its amazing!

[Tomatschek]

Hi Doom members!

I made the same movie with the same settings (prog&codec)
4 times...I only changed the resizer!

Xmpeg bressenham 130% AVI Size
Xmpeg mmx billinear 115% AVI Size
Xmpeg ssebicubic 100% AVI Size
m2a ogo bressenham 120% AVI Size

after compressing I compared the final AVI size.
If you set Xmpeg bicubic 100% the results shown
above are amazing. Could anybody explain these
differences in final AVI Size???? Why ist the more
detailed bicubic smaller than the less detailed
bressenham???

Please answer,
I would be very happy about any response =;-)

Thomas

[JimmyBarnes]

Originally posted by Tomatschek
Hi Doom members!

I made the same movie with the same settings (prog&codec)
4 times...I only changed the resizer!

Xmpeg bressenham 130% AVI Size
Xmpeg mmx billinear 115% AVI Size
Xmpeg ssebicubic 100% AVI Size
m2a ogo bressenham 120% AVI Size

What codec?

Did you use VDub or some other prog?

Did you use Fast Reprocess and and AVS video source file?

JB

[Acaila]

@JimmyBarnes
What codec?

Did you use VDub or some other prog?

Did you use Fast Reprocess and and AVS video source file?
-Since this is the DivX5 forum, I'd say he used DivX5.
-He already mentioned above that he used XMPEG and MPEG2AVI.
-N/A


@Tomatschek

My guess it the various resize filters discard/keep info that is invisible to the human eye. While one might seem blurrier it might have extra info in part you don't notice. Just pick the one that looks pleasing to you.
But since I never used those programs so I can only guess.

[Tomatschek]

Hi Acaila...

I agree with you - yes there must be more info written to disk.
But.....in every forum (in the Xmpeg develepers too) there is a
theorie:

Xmpeg bressenham 26 fps filesize 130%
Xmpeg bicubic 13 fps filesize 100%
(this are facts of my Duron 1200 System)

in all foren I got the same answer;
the brssenham resizer writes less data to disk (creates a smaller avi size)because it looks "worser" in quality than bicubic and this is the reason for it is sooo fast!!! (but you and I see it's wrong)

I have a own theorie (maybe it's wrong too)
the bicubis is soo slow because it searches more accurate if there is
compressable data in the frame or not. And this is the reason for the final size is smaller!!??

Could this be this way??
Thanx for response

Thomas

[Acaila]

Yup, that could be possible, but since I can't read the code I won't know for sure.

[theReal]

I've done a similar comparison with a small file (DV avi, hard to compress) with avisynth and virtualdub filters.

Filters were: bicubicresize in avisynth, used with the 3 Gknot presets soft, normal and sharp, bilinearresize in avisynth and precise bilinear resize in virtualdub.

The VDub precise bilinear resize won the filesize-competition (while all other filters produced files between 50 and 60MB, this filter made the file 45MB).
But, it is also not the worst-looking filter of the pack, it looks much better than bilinearresize in avisynth, and some might say it even beats the avisynth soft bicubic. So, you really have to test all filters for a)looks and b)compressibility before you can say what you want to use.

P.S.: I'd really recommend using Gknot/avisynth/VDub, because all of their filters beat the XMpeg filters by far. It's the resizing filters that make the difference (and why advanced encoders don't even consider using XMpeg)

[RadicalEd]

Well you guys have the right idea but its actually a bit simpler than that. Bicubic resize actually sharpens the image while it is resizing, bressenham softens. Because the image is sharper, there are higher frequencies, which are harder to code. A softer picture, on the other hand, has smooth frequencies that code well. Thus, a sharper image will always use up more bitrate than a soft one. The fact that bicubic sharpens is also most likely the reason why it takes longer.
Understand that most compression algorithms are based upon data being constant and redundant. Smoother frequencies of luma and chroma are similar and thus more redundant, easier to code. A huge frequency spectrum on the other hand will have much less redundancy and therefore will be harder to compress.

[theReal]

Not really... tomatschek's point was that bressenham should result in a smaller file than bicubic (because of what you said), but it doesn't.

Then I told tomatschek that I had a similar result with precise bilinear (VDub) vs. bilinearresize (avisynth).
The latter should result in a smaller file, because it softens more, but the clear winner was precise bilinear (it was sharper and smaller).

[RadicalEd]

ha.. misread, go figure (I need to be more careful). Hmm.. interesting situation indeed..

[JimmyBarnes]

Originally posted by theReal
I've done a similar comparison with a small file (DV avi, hard to compress) with avisynth and virtualdub filters.

Filters were: bicubicresize in avisynth, used with the 3 Gknot presets soft, normal and sharp, bilinearresize in avisynth and precise bilinear resize in virtualdub.

Is this published anywhere?

JB

[[NEX]]

I have take this from the VirtualDub manual.
· Bilinear uses a linear approximation in both directions using a 2x2 kernel. This is the second fastest mode, and produces very good results. It can reduce down to about 66% and enlarge up to about 8x without visible stairstepping. Most video cards with smooth video overlays have bilinear filtering in hardware. Bilinear filtering tends to produce a slightly blurry image and at very high magnification results in trapezoidal patterns. Because this mode only uses a 2x2 kernel, at high reduction it results in pixellation much like the nearest neighbor mode.


· Bicubic is the best choice for enlarging a picture, producing slightly sharper images at moderate levels and curves instead of trapezoids at high zooms. It is significantly slower than bilinear but produces much better at high magnfications such as 20x. Because bicubic mode treats detail better, very noisy video can appear worse under bicubic filtering than bilinear. Because bicubic mode only uses a 4x4 kernel, it too results in pixellation at high reduction.

As this says I have found that (to my eye) bicubic will allways give you better results because of the 4x4 kernel(that makes it takes so much time compared to the bilinear 2x2 kernel)

Taking in count that divX 5.xx is a high quality codec bicubic is the choice.

[JimmyBarnes]

Originally posted by [NEX]
I have take this from the VirtualDub manual.
· Bilinear uses a linear approximation in both directions using a 2x2 kernel. This is the second fastest mode, and produces very good results. It can reduce down to about 66%... Because this mode only uses a 2x2 kernel, at high reduction it results in pixellation much like the nearest neighbor mode.

Does 66 %% refer to linear or to area reduction?

JB

[theReal]

Is this published anywhere?
No. It was just a little test for myself in order to chose the right filter for a certain DV video.

Hey is this like at university? You don't have any publications and noone believes you? ;)

[JimmyBarnes]

Originally posted by theReal

Hey is this like at university? You don't have any publications and noone believes you? ;)

Umm.. I was more thinking, Published in a Doom9 forum thread somewhere :cool:

Terribly ignorant question coming up: What's a DV video?
(do you mean just a DVD?)

thanx
JB

[RadicalEd]

In order to regain my respectability in this thread, I'll answer this one (correctly :rolleyes: )

DV is the standard codec for digital video cameras. The stuff you record is stored on DV tapes and when captured saved as DV compression. Basically its just the master format for digital video cameras. It only compresses about 5:1, so it isnt very lossy and it is very high quality.

[Tomatschek]

Hi "the real" and everybody who's interested
============================================

since you did your comparison I worked on a second one...
now the resizer results (Duron1200 System;720x304YUY2):

m2a Bressenham 130% Size 22fps
xmpeg Bressenham 120% Size 22fps
xmpeg mmx bilinear 115% Size 17fps
xmpeg sse bicubic 100% Size 13fps
vdub precise bicubis 110% Size 11fps
vdub precise bilinear 100% Size 17fps
===now it comes=====now it comes=====
avisynth bicubic(0,0.80) 110% Size 22fps
avisynth simpleresize 100% Size 26fps

the same system,
xmpeg ssebicubic creates a simmilar quality to
avisynth simpleresize but avisynth did it in
half the time!!!

Last but not least a little question:
The maker of the simpleresize.dll wrote that it is
a two tap anistrophic filter. What does this mean?
Is it like bicubic or near bilinear or a third method?

Thanks for any responce.

Thomas
carpe diem

[Acaila]

Is it like bicubic or near bilinear or a third method?
SimpleResize is bilinear. But without the blurring that avisynth's native BilinearResize() does.