automatic controlpoints generation

[nobody]

As I work mostly with a fisheye lens (a full frame in my case).

If I try a full automatic creation of control points, matching ok, except at the edges (ie where image distortion is the worst)

I notice that control point matching is better on the central zone of photos, thus, I though that doing a pre-stitch (using an autopano first pass and an optimisation of y,p,r,v and b) should be helpful to autopano to find and match some control points close to the edges of these pre-stitched images.

Then, these control points locations have to be remapped to the orginal images, and the optimisation process can be done one more time. May it work better ?

I hope I was clear enough, and that the idea is something great.

[nobody (nobody-users)]

Logged In: NO

You might be able to get that result through a series of manual steps:

1) generate (automatic) control points
2) optimize
3) for each pair of overlapping images, select them in the control points tab and press 'g'
3b) fine tune control points and remove problematic points
4) re-optimize
4b) remove outliers and re-optimize again

I use that process with regular images and selecting 2-3 control points by hand in step 1 (and ending with a photometric optimasation pass), so it might or might not work with fisheye.

I even thought about writing a second assistant mode, replacing steps 1 and 2 with a pano preview where the images can be aligned by dragging. But I haven't come around to that, because I have not yet gotten annoyed enough about having to revisit every image in step 3 :)

[Brent Townshend (bst)]

Logged In: YES
user_id=1934048
Originator: NO

I'm doing exactly what you describe with my fisheye panoramas. I do a first pass rough alignment and then run nona (using the -c to save the coordinate files) to create a set of mapped images. I then run autopano on these and get lots of good control points. I then map the control points back to the original domain using the the coordinate files and some MATLAB scripts I wrote and can start again with a much better set of control points. Works for me, but my workflow is such that it would not be very easy for anyone to use my scripts. I suppose you could also just realign the first pass aligned images using the new control points, though that would lose some performance.

It seems to me that this kind of operation should be integrated into Hugin, perhaps on a patch-by-patch basis to allow better alignment. Though I think that getting the whole process smooth would require integrating autopano into Hugin actual code rather than relying on external programs.

[Pablo d'Angelo (pablo.dangelo)]

Logged In: YES
user_id=30308
Originator: NO

There are several things I have planned to cope with the distortion:

1. running autopano on images in stereographic projection. This is a conformant projection and preserves local shape at the expense of different scaling. I tested some images, and points where also found near the corners. I planned to integrate this approach into hugin after the next release.

2. Use a control point generator aware of the mapping (mostly during extraction of the descriptor).

[tmodes (tmodes)]

This is implemented in autopano-sift-c and cpfind