Which resizing filter do people prefer?

[kinneera]

In reading through the filter descriptions for AVISynth, I noticed that it says the bilinear resize filter will usually produce equivalent or even better quality than the bicubic filter(s) when resizing to a smaller size (ie, DVD to VCD/SVCD). Yet, the majority of AVISynth scripts I've come across use bicubic. I was wondering what other people prefer, and possibly an explanation why?

[Truman]

Bicubic sharpens the image, I prefer a sharp image.
If I resize both horizontal and vertical, (well, usually I need both) I use bicubic b=0, c=0.5. Higher c values may result in a too sharp image ("ghoosts" appear)
If I resize horizontal only, I use bilinear.

[kinneera]

Any thoughts on FitCD's suggestions to optimize for CCE?

<TABLE BORDER=0 ALIGN=CENTER WIDTH=85%><TR><TD><font size=-1>Quote:</font><HR size=1 color=black></TD></TR><TR><TD><FONT SIZE=-1><BLOCKQUOTE>
BicubicResize(352,478,0,0.6,7,0,706,480)
AddBorders(0,1,0,1)
</BLOCKQUOTE></FONT></TD></TR><TR><TD><HR size=1 color=black></TD></TR></TABLE>

[Truman]

FitCD's "optimize for cce" option refers to macroblock optimization. It calculates how to resize, if you have top and buttom borders. If you use cce, add 17 pixel borders instead of 16 (resp. 129 instead of 128 etc.). There is no difference between 0 and 1.
I assume, you want to resize to half D1, so I suggest eigher to crop a little more:

BicubicResize(320,446,0,0.6,15,0,690,480)
AddBorders(16,17,16,17)

or resize to a smaller frame

BicubicResize(320,436,0,0.6,7,0,706,480)
AddBorders(16,22,16,22)

Always add 16 pixel borders left and right. There is no reason to encode, what you can not see (TV overscan).

[kinneera]

Sorry to be a bother, but I'm looking for some clarification as to the syntax - your resize statements don't seem to match the Avisynth reference. Specifically, what are the "15,0" and the "690,480" and the "7,0" and "706,480", and why was the 15/7 and 690/706 chosen?

Also, are any other changes necessary if the source is 16:9, not letterboxed and needs to be converted to 4:3 letterboxed?

[Shabubu]

The one problem with the sharp image is that it won't compress as well, also the bilinear filter will help for noisy sources. I bascially test each with the individual vids I'm doing and see which looks better for each vid. Technically (according to the paint programs) you should use bilinear for down and bicubic for up.

[Truman]

Of course the script match the Avisynth reference.
BicubicResize(target width, target height[,b,c],crop source left, crop source top, source width, source height)
Both scrips above are for 16:9 (not letterboxed) to 4:3 (letterboxed) conversions, target resolution 352x480 (half D-1).
Yesterday a couple of topics about "TV overscan" came up. I use to add borders to the overscan area, why should I encode a bunch of macroblocks, that I can not see at TV?
A macroblock is 16x16 pixels, that's why 16 pixel borders left and right (recall macroblock optimization). Check the FitCD readme for more information or this great, great german site:
http://www.uni-kassel.de/~eckhardm/hq.htm
The google translation is quite strange, but have a look.