[ptx] Need control points matched across multiple images,
not just pairs
Littlefields - Rik, Janis, Kyle & Peter
rj.littlefield at computer.org
Mon May 3 17:53:45 BST 2004
In this email I talk about seams and control points.
I have changed the subject line to reflect the most
Pablo d'Angelo wrote:
>>1. The PTStitcher masks often seem badly placed.
>>I would expect seams to be at the middle of overlap between
>>two images, so that every pixel in the output image comes
>>from the input image in which it is most central. PTStitcher
>>seems to do something different, so I was hoping to either
>>understand why I am thinking wrong, or fix the code.
>Hmm, the next version of nona will place the masks based on the nearest
>feature transform, as implemented in enblend.
I don't know what "nearest feature transform" is, and I have
not worked with enblend, so I can't follow what you wrote.
However, I would like to share with you the strategy that
I use to make panoramas. I think it is different from what
others do, and I see a problem that I have not heard others
Background is that I often shoot handheld panos, for example
while backpacking in the wilderness. The most interesting
ones always include some nearby objects as well as distant
ones, and of course the nearby objects are vulnerable to
parallax error as well as movement due to wind, etc.
Even the clouds typically move quite a bit between frames.
The strategy that I have evolved to deal with these
difficulties goes like this:
1. Roughly position source images into proper alignment.
(I use the preview window of PTGui.)
2. Estimate where the stitcher will put seams.
3. Pick high quality control points near the seams.
"High quality" means stable between frames
and placed on visually important objects,
places where a mismatch would be obvious.
4. Pick additional control points needed to make
step 6 work well.
5. Optimize parameters.
6. Render with morphing, so that all control points
line up exactly. (It helps here if there are control
points also away from the seam, to make the
triangulation more uniform.)
7. Adjust layer masks to minimize any remaining
This strategy generally works quite well. However...
There is one fundamental problem in the design
of Panorama Tools and supporting gui's that makes
it less effective than it could be.
The problem is that control points come in pairs.
Let me emphasize a bit here.
*** Control point PAIRS are the wrong model. ***
The right model, from my standpoint, is that there
are control points *in the world*. A world control
point often appears in several images, and when
it does, I want all of those images to line up, at
that world control point, in the rendered image.
But Panorama Tools and the gui's that I know about
provide me no way to make that happen. I will not
attempt to detail here what goes wrong. Suffice it to say
that the morpher can generate nasty rendering
artifacts if one image has two control points that
are close together and are tied to separate other images.
So I am forced to *avoid* using the same world
feature as control in multiple image pairs, when using
What I would really like is:
1. (Most important) Support for *world* control points
in the data models, user interfaces, and underlying software.
2. (In comparison, nothing else matters here.)
I will not take space here to talk about ways in which
this capability might be patched into Panorama Tools.
The important thing is the idea that I want
control points that are to be matched across
multiple images, not just pairs of images.
Is this already included in the newer software,
and I just don't know about it?
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the ptX