[ptx] panosifter problems & questions

[Pablo d'Angelo]

On Thu, 12 Feb 2004, scholnik at radar.nrl.navy.mil wrote:

> I've been playing around with panosifter since it appeared in CVS
> and had some success, primarily with narrow-angle (36-108mm equiv)
> shots near the horizon.  I just did a 1+ row 360 pano with 45 total
> 1200x1600 shots, and it found matches for most of the overlaps.  Now
> I've got a new camera with a 19mm (equivalent) wide angle lens and I'm
> finding things a lot more difficult.  I just did a 2 row x 9 portrait
> shots/row (about 33% overlap) + zenith and panosifter is not finding
> many matching photos, no matter the subsampling parameters I use.
> When I go into hugin to add points myself I'm finding that
> auto-estimate and fine-tune aren't working so well either.  I'm
> assuming this is because hugin and panosifter don't de-warp
> (pre-warp?) the images and are looking for pure translations?

hugin uses simple correlation, and will break. the sift features are
invariant some perspective distortion, but not truely projective invariant.
While it didn't work as good as with 50° hfov images, I had some success
with circular fisheye images.

To transform into a common reference frame, one needs to know the
transformation, which we are trying to calculate ;)

> Are
> there any plans for using rough yaw/pitch/roll/fov/abcde estimates to
> project the current two images to a common reference frame for
> control-point picking?

At some point I might introduce that. but until I get a fisheye or another
wide angle lens this has low priority for myself. Unfortunately I'm rather
low on time, and can't implement all the ideas floating around
immediately. So I concentrate on stuff that is useful for me. Feel free to
contribute :)

> 1. panosifter tries to read focal length and sensor size from the exif
>    data, but when it can't find the sensor size it throws an error and
>    assumes 24mm x 36mm.  My camera stores the true focal length but it
>    has a cropping factor of about 5, so assuming 24mm x 36mm is really
>    wrong.  I do use the -v parameter but I think it is being ignored.

haven't tested -v that myself yet. (panosifter is just a quick hack,
really).

> 3. The recent CVS panosifter seems to be broken; it decides that every
> 	image matches every other image, and uses the same set of points
> 	for every image pair.

Yep, I'm trying to use our own sift implementation (which is broken at the
moment). Hmm, maybe I wasn't careful enough while changing the stuff.

> I can provide images or debugging output upon request.

since this is all temporary stuff, wait until the next version is available,
or use the older version of panosifter.

ciao
  Pablo
-- 
http://wurm.wohnheim.uni-ulm.de/~redman/
Please use PGP