proj-imim:
Re: Number of Control Points, was Re:
Peter Murphy 2001-Jan-05 19:57:26
I agree with most of this.
I think also that large correction factors are generally a sign of
poor optimisation.
I think that for carefully shot horizontal sequences 3 control points
per
image seam is fine.
I usually look at the discrepancies and remove the worst one or two
point
pairs-
point marking introduces its own errors.
I think it is important to keep in mind the possibility that a good
stitch
might not be physically accurate but generally small discrepancies and
small correction factors are a sign of a physically accurate calibration
- and generally this corresponds with good stitching.
t factor lines make it much more certain that ones answer is
physically accurate. I would usually use at least a couple of t1 factor
lines for at least one image (generally the first) for any
sequence where scene verticals were visible. In this case you
must optimize r0 and p0 and not set them to zero as one
normally would.
I think handheld and tilted sequences are a special case.
Here I think it is most important that one have a good idea of
v to start with and stick to it and not optimize it at all or only put
it in
at the end. Similarly I would usually stick to predetermined
distortion factors til I felt I had y, p and r's into the ballpark
With tilted/handheld sequences the use of the t1 (and occasionally when
the horizon is visible t2) flags is most helpful.
Usually for these cases I use 3 regular point pairs and 2 or 3 same
image t1
lines for each image in these sequences - thus six tilted images would
have 18 regular point pair lines and 12 to 18 t1 lines.
In difficult cases (or where you have made an error preparing
the script) the answer usually is to build up the panorama from
the initial pair. Here if there is going to be a problem it is usually
at the end when you are trying to optimise the final "wrap"
pair.
You should generally choose distant points.
You should generally add the "morph to fit" flag.
I would be better if you could get a=a0 etc to work
like you can v=v0.
I think it is helpful to have a general idea of how much
the CCD is decentered with digital cameras and apply
this cropping to the images before you start - the only easy
way to do this is to put a circular fisheye on the camera
and note the offcenteredness.
If you do up and down shots they should not be included
in the main optimisation. They might be best optimised
by calibrating them to the stitched equirectangular image
resulting from the first lot of calibration from the
main sequence.
cheers
Peter Murphy
John Blommers wrote:
> Let me share the results of a little experiment I just did to try to
> understand all this business about how many control points are needed.
> I have 4 fish-eye images and this PTPicker script: p f1 w2400 h1200
> v360 u30 n"QTVR"
>
> i f2 w1152 h1152 y0 p0 r0 v183 n"A01.jpg" a0 b0.1 c0 X0 Y0 Z0
> i f2 w1152 h1152 y90 p0 r0 v=0 n"A02.jpg" a=0 b=0 c=0 X1 Y0 Z0
> i f2 w1152 h1152 y180 p0 r0 v=0 n"A03.jpg" a=0 b=0 c=0 X2 Y0 Z0
> i f2 w1152 h1152 y270 p0 r0 v=0 n"A04.jpg" a=0 b=0 c=0 X3 Y0 Z0
> v v0 a0 b0 c0
> v y1 p1 r1
> v y2 p2 r2v y3 p3 r3 There are 4 images to stitch and 13 variables to
> optimize. I started out by picking lots of extra points, then deleted
> more and more until I had only two points for each image seam. The
> results are most interesting: Number of control pointsAverage Error
> after ~2000 iterations243.8164.1124.408 points2.0 The last experiment
> gave the best stitched result! The optimized v value for the fish-eye
> lens came out to 183.2 which is what the client said it should be. But
> in the first trial the optimiser came up with 184.3, slightly higher.
> More importantly, the last trial's values for the lens correction
> parameters (a b c) were a lot smaller than the first trial. My
> experience is that the smaller the lens correction the less distortion
> in the final stitch. Perhaps "less is more" works here too. Let me
> also point out that each control point represents 4 independent
> variables since the point is present in 2 images and each point has an
> independent (x,y) coordinate. This little fact is apparent when you
> look at the definition of a control point: c n0 N1 x768 y262 X363
> Y258 The lesson for me is that maybe I can get by with a lot less
> control points than I thought. Perhaps other list readers will share
> their experiences in this matter, - John