[ptx] suggestions for stitching aerial photos

[Rik Littlefield]

Rob,

Sorry, but your problem is not as simple as finding unknown lens
parameters.

These aerial photographs were taken from different positions as the
airplane flew along.  As a result, variations in elevation have turned
into variations in horizontal displacement -- details are moved
relatively left or right depending on whether their elevation was high
or low.  For stereography, this is useful information; for stitching
it is uncorrectable distortion.

Take a close look at the right hand bridge in images ca74_012.jpg and
ca74_013.jpg, and you will see what I mean.  In 013, we are
essentially looking straight in line with the bridge, and it appears
to be a single line across the water.  In 012, we are looking sideways
under the bridge, and the supports are clearly visible.

These two pictures actually make a great stereo pair.  As I write
this, I am matching the upper 300x300 pixels of 012 (right eye view)
against 013 (left eye view).  I can clearly see the large T-shaped
building sticking up from the landscape, bluffs along the river near
the long bridge, and a hillside angling down to the left from the
T-shaped building.  Pretty cool!

(If you're not good at viewing stereo, then pull both pictures into
Photoshop, make a 2-layer stack of them, align the layers so the
T-shaped building is in the same place, then click the top layer off
and on repeatedly.  It will look like the terrain is rotating back and
forth under you.  Even if you are good at seeing stereo, you will
probably be surprised at how strong and intuitive the rotation effect
is.  But I digress...)

It may also be that the pictures have some systematic distortions in
them and between them.  However, I looked for those by letting the
optimizer play with various combinations of a/b/c, d/e, and fov,
globally and for each image separately, and I was not able to make
much improvement.  It is still conceivable that the pictures have
distortions that Panorama Tools cannot handle, such as keystoning due
to the camera not pointing straight down.

But I am inclined to think that your biggest problems are lateral
shift due to stereo effects, and I do not know of any practical way to
get rid of those.

In the absence of stereo effects, the strategy that you used would
have worked pretty well -- just shrink the fov as small as your GUI
will conveniently handle.  If the pictures might have been taken from
different altitudes, then optimize fov independently for all but one
image.

--Rik

Rob Park wrote:

 >Hi everybody,
 >
 >I recently discovered this website:
 >
 >http://airphotos.nrcan.gc.ca/photos101/edmonton_e.php
 >
 >Which contains aerial photos of my city from 1924 and 1950. Naturally,
 >I decided to stitch them together, and I came up with this:
 >
 >http://rbpark.ath.cx/panos/1950.jpg
 >http://rbpark.ath.cx/panos/1924.jpg
 >
 >Now, this was my first time stitching images that weren't produced by
 >my own camera, and didn't have EXIF data. I wasn't sure what value to
 >put in for the FOV, but I noticed that I got less distortion as I
 >decreased the value. The panoramas above were produced by setting the
 >FOV to 2 degrees. It came out pretty good, but there are still some
 >stitch lines. Does anybody have any tips for how to guess unknown lens
 >values?