While I understand as a general matter that smaller control point
distances (as reported by the optimizer) imply higher quality stitched
output, imagine my surprise when I got a message from the optimizer
that the maximum control point distance for the first image in the
above gallery (the pin oak leaf) was "too good to be true". In this
case it was 0.726523 pixels.
Given that control point coordinates are stated in pixels, how do you
end up with a maximum control point distance that is less than 1 pixel?
> Given that control point coordinates are stated in pixels, how do you
> end up with a maximum control point distance that is less than 1 pixel?
I believe the control point distance is how much a certain point on
one image varies in relationship to the same point on next one that it
overlaps. in your case they were slightly less than 1 pixel off, which
is very good. it means the stitcher had an easy time of selecting the
exact same point on the two overlapping images.
"Too good to be true" is often the result of poorly distributed
control points and/or too few points. It's easy for the optimizer to
align a few points in the middle of the overlap areas, but this does
not ensure that the image alignment will be equally good in those
areas devoid of points.
John
On Feb 10, 12:18 am, Hambagahle <lionel_d...@yahoo.com> wrote:
> While I understand as a general matter that smaller control point
> distances (as reported by the optimizer) imply higher quality stitched
> output, imagine my surprise when I got a message from the optimizer
> that the maximum control point distance for the first image in the
> above gallery (the pin oak leaf) was "too good to be true". In this
> case it was 0.726523 pixels.
> Given that control point coordinates are stated in pixels, how do you
> end up with a maximum control point distance that is less than 1 pixel?
PTGui calculates the control point distance by projecting the control point from the second image into the first image. It's possible that the projected control point ends up somewhere between two pixels so you can have a distance that is not a whole number, or less than 1 pixel.
As John explained 'too good to be true' often means just that, but if you have control points well spread over the entire overlap area then it just means that the images are aligned extremely well.
> While I understand as a general matter that smaller control point > distances (as reported by the optimizer) imply higher quality stitched > output, imagine my surprise when I got a message from the optimizer > that the maximum control point distance for the first image in the > above gallery (the pin oak leaf) was "too good to be true". In this > case it was 0.726523 pixels.
> Given that control point coordinates are stated in pixels, how do you > end up with a maximum control point distance that is less than 1 pixel?
Sometime around 9/2/11 (at 23:18 -0800) John Houghton said:
>It's easy for the optimizer to >align a few points in the middle of the overlap areas, but this does >not ensure that the image alignment will be equally good in those >areas devoid of points.
Absolutely. The trick is to use a conservative number of points and make sure those are spread relatively well across the whole overlap area. Clustering points together won't help with real accuracy, that's no more effective than hanging a picture by putting a dozen nails into one side.
I know people sometimes put in loads of points because they're not confident about individual point accuracy. That just averages out the uncertainty to some uncertain amount, it doesn't guarantee any improvements. One well-placed point is far more use than a collection of imprecise ones, so spend time on those.
With my fisheye shots I aim for four points per image pair, spread right down the overlap area. I often get very low single-figure values straight off, sometimes in the zero-point range, and I've seen "too good to be true" a handful of times. (And yes, the stitch has been damn perfect in those cases, so I think the cynical message was misplaced. :-)
I've often wondered if these numbers were actually residuals from a
least squares analysis - is this possibly true? (I use least squares
in my day job; land surveying :)
Kelly
On Feb 10, 4:50 am, PTGui Support <supp...@ptgui.com> wrote:
> PTGui calculates the control point distance by projecting the control
> point from the second image into the first image. It's possible that the
> projected control point ends up somewhere between two pixels so you can
> have a distance that is not a whole number, or less than 1 pixel.
> As John explained 'too good to be true' often means just that, but if
> you have control points well spread over the entire overlap area then it
> just means that the images are aligned extremely well.
> > While I understand as a general matter that smaller control point
> > distances (as reported by the optimizer) imply higher quality stitched
> > output, imagine my surprise when I got a message from the optimizer
> > that the maximum control point distance for the first image in the
> > above gallery (the pin oak leaf) was "too good to be true". In this
> > case it was 0.726523 pixels.
> > Given that control point coordinates are stated in pixels, how do you
> > end up with a maximum control point distance that is less than 1 pixel?
> I've often wondered if these numbers were actually residuals from a > least squares analysis - is this possibly true? (I use least squares > in my day job; land surveying :)
> Kelly
> On Feb 10, 4:50 am, PTGui Support<supp...@ptgui.com> wrote: >> Hi Lionel,
>> PTGui calculates the control point distance by projecting the control >> point from the second image into the first image. It's possible that the >> projected control point ends up somewhere between two pixels so you can >> have a distance that is not a whole number, or less than 1 pixel.
>> As John explained 'too good to be true' often means just that, but if >> you have control points well spread over the entire overlap area then it >> just means that the images are aligned extremely well.
>>> While I understand as a general matter that smaller control point >>> distances (as reported by the optimizer) imply higher quality stitched >>> output, imagine my surprise when I got a message from the optimizer >>> that the maximum control point distance for the first image in the >>> above gallery (the pin oak leaf) was "too good to be true". In this >>> case it was 0.726523 pixels.
>>> Given that control point coordinates are stated in pixels, how do you >>> end up with a maximum control point distance that is less than 1 pixel?
The maximum control point distance can be used to tell you how you are
doing. You still have to examine the completed panorama for stitching
errors. Fisheye lenses (like the Sigma 8mm) tend to have control
points bunched up along the horizon. After adding control points at
the Zenith and Nadir and optimizing you get "Too good to be true" you
still have to examine the panorama. If there are no stitching errors
and you have a maximum distance of 0.x, you have a set of images that
are aligned very well. It is more about looking at the finished
panorama than relying on a set of numbers.
Well, thanks everyone, that's very interesting. The insights in all
the replies are very helpful. Luckily I do inspect my stitched images
thoroughly; these passed my test, and I'm really pleased with them.
Sometime around 10/2/11 (at 07:33 -0800) DennisS said:
>Fisheye lenses (like the Sigma 8mm) tend to have control >points bunched up along the horizon.
That's if you ask PTGui to create control points for you automatically, right? I rarely do that these days; I find that creating a few points manually gives excellent results, and it doesn't take any longer than generating them automatically and cleaning up the results.
>It is more about looking at the finished >panorama than relying on a set of numbers.
Definitely. The numbers help a lot, but always look carefully at test stitches and the final output.
