On September 8, 2010 02:02:40 am Tduell wrote:
> I have spent a bit of time lately trying to get a good understanding
> of how to work with mosaic mode, all aimed at putting together a
> tutorial for the 2010.2 release.
thanks for sharing your experience. I built on it [0].
> (1) Set the default control point detector to "pre-aligned panorama",
> and ensure that the advanced selection "only work on image pairs
> without control points" is selected.
Done away with it. Use manual CPs instead, is faster.
> (7) Fast Preview window, layout, mosaic. Drag the new image roughly
> into position.
Use the numeric input in the Images tab instead, and you won't be limited by
the drag/move functionality.
> The above process is clearly tedious, particularly when a project has
> a lot of images, and not a desirable solution, but at this stage it
> seems to to needed to get a result with some projects.
The problem is that dragging will move all images connected by CPs, so you
must do the dragging before adding CPs or find an alternate way to get the
image into position before optimizing.
Because optimization is heavily dependent from the initial position of the
image, dragging it (or setting it there by entering a reasonable numerical
value in the Images tab) is not an option. What Hugin needs is an extra
switch to the dragging functionality in the Fast Preview window: drag group
of connected images (as exist now) and drag individual image (new).
> Bruno's project has been interesting. There have been situations where
> it has all been going quite good until adding in the last image, then
> it goes completely awry, and no amount of recovery tricks seem to
> work.
once things go awry it is very difficult to recover. I found that the
starting positions are much more important in determining the outcome than in
traditional panoramas on the sphere.
> If anyone would like to try to track down the problems in the code
> that are causing these issues, images #178 and #180 in Bruno's project
> might be good test case.
I have not played with Bruno's project and have arrived at similar conclusions
as you. There is room for improvement.
Yuv
[0] http://panospace.wordpress.com/2010/09/19/linear-panoramas-mosaic-
tutorial/
If you select Identify in the Preview window and hover your mouse over
either the row of image number buttons or the preview image itself you
can see the individual photos.
Hugin can't show you where the seams will be because enblend will do
it's own thing and this can't be accurately predicted. A long time ago
before we had enblend, the Hugin preview did actually draw simple seam
lines, this code is still there but disabled, there is probably
something useful that can be done with it.
--
Bruno
On September 20, 2010 09:38:48 am Oskar Sander wrote:
> Nice walk though!
thanks,
> * Is there something inherent in CP detection that is thrown of by linear
> mosaics? My experience is too that it doesn't work well, but i tho
> thought it was my low-contrast low-light and monochrome application (very
> dark underwater scenes) that was throwing the search off.
depth (pun intended). The CP detectors will detect features that look the
same, but they still can't tell if all of those features are on one plane.
> - One experience that I have form manual CP detection is that IF your
> mosaic subject is not a flat plane, you need to analyze the images and make
> sure that the CP are selected in the image plane you want
exactly that. My experience with building fronts is that even small relief
makes a big difference.
> When you are laying the images out manually, are you
> just picking a number?
I help myself visually with the feedback of the fast preview.
> The X,Y,Z parameters are a scaling factor of the
> radius to the unit panosphere I understand.
the unit is relative. The way I understand it is that with X you shift the
center of the panosphere left and right. With Y you shift it up and down.
with Z you shift it forth and back ("zoom"). Each image is on its own
panosphere that gets projected on the mosaic surface. Knowing this, playing
with the numerical values in the Images tab becomes a "fly by wire" exercise
where you zero in on an approximate position, before optimizing it with the
optimizer.
> Maybe a simple calculator
> button next to the X, Y X fields to place the image a whole number of image
> lengths away from the anchor could be useful.
The image length is constant on the mosaic only if you shoot from the same
distance and angle - a special case. I don't see it as being helpful to me.
I rather have some sort of 3D joystick/trackball. one joystick would do X/Y
the other would do Z/roll and the trackball would do pitch/yaw.
> * I like your in-line lens calibration in the work-flow, I have to try
> that to se if I get better results!
It's a fortunate coincidence that I have the necessary features in the
picture. It's really much better to work with calibrated lens profiles in the
first place, but it has been so long since I've been serious about pano
stitching that I have not bothered to calibrate this lens I bought about a
year ago (Sigma 15mm fisheye).
> Another question that I think I asked before but can't find the answer to
> is: When you have generated a mosaic project, is it possible to generate
> an output where all image seams are highlighted?
If my memory does not betray me, there was a debugging version of enblend that
would do just that, but I can't find any references to it right now.
Given the extreme sensitivity to depth/perspective, it is anyway better to
place the seams manually, using masks. That's the most time consuming part of
the work. Fiddle, fiddle, until it looks acceptable.
Yuv (still fiddling)
The example I'm providing is an extreme view. I did some "rough"
optimization/layout in Gimp. I suppose this is what I was expecting.
Using Shift, Perspective, & Scaling in Gimp, I was able to get close. I
think some type of curve warping of the photo would get one closer. Add
either manual masking or aggressive enblend masking, and one might be
able to combine the two photos. They are from different angles,
different distance, different tilt, etc. Everything that could be wrong
is here.
I would expect to be able to add manual CP or use a different kind of
mosaic mode CP detector (some modified form of geodaisy) to find common
control points. There are some "constants" or common control point
possibilities in the photos, such as the picture frame, stone corners,
window corners, wall chair, and to some degree, door frame. If one was
able to connect the dots with these constants and then warp each photo
to fit, mosaic mode work work well. The constants are "flat" objects
not subject to change from perspective, shift, tilt, etc.
Attached are the files. Interested in feedback and possibilities.