[ptx] New features in current hugin CVS

[Pablo d'Angelo]

Hi!

I've got internet at home again...
Somehow it wasn't that bad to be without inet connection.. less distraction
from writing code, and creating panoramic images. During processing some of
my images from the holiday, I noticed that for a really good fit, manually
placed control points a quite valuable, especially with moving subjects.
Therefore I have worked on the fine tune and control point setting a bit.

Several new features:

1. fine tune can, optionally, search for rotated patches.
   This is a lot slower, because of my inefficient correlation routine,
   but works quite well. Enable it with the following parameters
   in the config file/registry.

   CPEditorPanel/RotationSearch = 1  # enable

   And tweaked with the following parameters:
   RotationStartAngle  -40    # start search with angle (default: -40 deg)
   RotationStopAngle    40    # stop search at angle (default 40 deg)
   RotationSteps        10    # number of steps between Start and Stop Angle 

   I've used these values with my 50 deg hfov images, and the allowed
   good matching even in the corners of shots that were tilted down heavily.
   
   I suspect that this feature is also very useful for fisheye images,
   but I'm not sure which angle intervall should be searched for them.

2. Fixed crash on remove image
3. fixed bug that prevented pano creation, if the output file was specified
   with extension (windows only bug)

4. after a "finetune all", the control point errors are replaced by the
   normalized cross correlation of the control point patches. Points with a
   low correlation can be spotted easily then. (A value from 0.75..0.8 to 1 is
   usually good, lower areas indicate a problems). (finetune all will also
   use the rotation search, if it is enabled)


Features to come / partly written

1. I'm about to allow multiselection in the CP list, so that
   multiple control points can be selected and removed based on their error
   distance easily.
   Hmm, any idea for a useful default threshold? Is there a good threshold
   based on statistics of the control point error? (I'm not sure if this
   feature is really a good way to get rid of bad control points...)
   
   Kind of manual AP-Clean.

2. Additional, magnified and perspective corrected display for
   the area around the control points. This should be very useful
   to set control points on heavily distored images (fisheye), but
   it's currently disabled, because there is one critical bug left.

   Maybe somebody of the other developers (Rik or JD) can help me out here:

   1. I set the two images to r,p,y = 0, and compute (using the inverse
      transform), the spherical coordinates (cp_yaw, cp_pitch) of the
      control point.

   2. I'd like to rotate the image, so that the control point is placed
      at the center of the pano (r,p,y=0).

      Simply using -cp_yaw and -cp_pitch for y and p doesn't work, because
      it looks like y,p are not defined as rotations about fixed angles,
      but as euler angles.

      However, I haven't come up yet with the proper way to calculate the
      r,p,y values. Any pointers?

   3. remap these images using these parameters, and display result
      in a separate window

  Except for 2. everything is already written.

  Using this information, the fine tune could be done on perspective
  corrected and correctly scaled patches, which should also improve
  the fine tune accuracy a lot, I hope. Combined with the subpixel
  correlation, it should be possible to get really accurate fine tuned
  control points (with a error < 0.1..0.2 pixels or so, I hope).

PS. sorry, no windows exe yet.

ciao
  Pablo