Re: [PanoToolsNG] Re: Some more tests on Optimal Cubefaces.
- Erik Krause wrote, On 01.11.2008 15:20:
> Am Saturday, November 01, 2008 um 13:17 schrieb Hans Nyberg:The "bi" in bicubic comes from the fact that the filter is used 2 times,
>> CubicConverter uses Bicubic which seems to be called Cubic in wiki.
>> Why not use the Bicubic definition if it is the same. I never heard
>> anyone call it just cubic. Photoshop today has 3 Bicubic versions.
> I don't know whether it is really the same. However, apparently the
> algorithm was called cubic in the originating paper:
one time in X and one time in Y direction. If you use such a filter on a
CT scan they are often called "Tricubic". As Panotools and Pano2VR
provide more then one filter the "bi" is kind of redundant as it should
be clear that they are used in both directions. Otherwise the filters
should have been called "Bigaussian", "Bilanczos", "Bimitchell"...
>> From what I understand Pano2VR uses Mitchell as default but Mitchell softens theI also use a kernel size of width/4 for the conversion. This is between
>> cubefaces slightly compared to CubiConverter. They need 0,4 100% unsharp mask to get
>> back to the same quality if you do an conversion with editing cubefaces and convert back
>> to equirectangular. This does not make sense after reading your comment as Mitchell
>> should have the same effect as Cubic.
>> Lazcos 3 gives normally same quality as CubicConverter but I found that I had to get up
>> to Blackman/sinc Filter to get the same when doing downsizing.
> Interesting. I think Thomas should clarify...
the two extremes of width/Pi and width/(Pi*sqtr(2)) (see my previous
post about the cube face sizes) and is a compromise between "too soft"
and aliasing. If you don't mind aliasing you can turn of the "dynamic
kernel" in the preferences and you should get a sharper image.
- Erik Krause wrote, On 01.11.2008 14:52 Uhr:
> BTW.: While photoshop creats nice small JPEGs at good quality itThe reason for these artifacts is chroma subsampling
> might be no good idea to use it for cubefaces stripes. In any stripe
> there are at least two joints where the single cubefaces don't fit.
> At these joints the jpeg compression creates artifacts that are later
> visible in the resulting panorama. See
> http://www.photopla.net/wwp0703/stripes.php for details.
> If I remember correctly Pano2VR and Pano2QTVR create stripes using a
> different compression scheme near the edges, which avoids these
In Pano2VR and Pano2QTVR the subsampling is turned off for stripes.
Please note that the tile size must be a multiple of 16 to make this work.
In Photoshop different levels of subsampling are used depending of the
quality setting. To be on the safe side with CS2 you need to use at
least Quality 7 in "Save as..." and Quality 51 in "Save for Web". For
more information you may also have a look at:
- Am Saturday, November 01, 2008 um 17:55 schrieb Philipp B. Koch:
> Thanks a lot, Erik! I've tried both Lanczos3 and Mitchell with DOSUPHow much do you downsample? I did some tests some time ago and found
> like you proposed. The visual difference is surely worth the (much)
> longer processing time it takes compared to spline64...
no big difference for a 4000x2000 to cubefaces 1200 remapping tasks.
I stopped testing, since the old fixed kernel size interpolators and
the anti-alaising ones are not comparable. The kernel size (and hence
the execution time) highly depends on whether downsampling or
upsampling, and they increase for downsampling:
Offenburger Str. 33
- Erik Krause schrieb:
> Am Saturday, November 01, 2008 um 17:55 schrieb Philipp B. Koch:I've made some tests with an equirectangular 4742 x 2371, downsampling
>> Thanks a lot, Erik! I've tried both Lanczos3 and Mitchell with DOSUP
>> like you proposed. The visual difference is surely worth the (much)
>> longer processing time it takes compared to spline64...
> How much do you downsample? I did some tests some time ago and found
> no big difference for a 4000x2000 to cubefaces 1200 remapping tasks.
> -> http://www.panotools.org/mailarchive/msg/41713#msg41713
> I stopped testing, since the old fixed kernel size interpolators and
> the anti-alaising ones are not comparable. The kernel size (and hence
> the execution time) highly depends on whether downsampling or
> upsampling, and they increase for downsampling:
it to 1052 cubes (=~ / 4.5) with spline64, Mitchell and Lanczos3.
I did not count the time for each task, but both Mitchell and Lanczos3
took well three times as long as spline64, I'd estimate.