As shown in the APG info pane, global RMS is very poor (10.7) but, as they are 700 images, 3,869 links and 151,634 CPs, sorting out the wrong ones is not an easy task!

One solution could be to click the CP editor icon, ignore the CP editor window and zoom the preview toward the places where you found stitching errors and delete useless links (4 links per source image is enough) or select only the worst ones (one after another!) and open the CP editor window.

Another method could be to use the Hide links below this value option and to delete or edit those worst links.

Yet another method: click on the settings icon in the CP editor window and "play" with Optimization stage options like Remove links over this error, Keep only the N best control points and Keep only control points below this error. This is something I often do for ...a dozen of source images, for 700 source images I'm not sure this is feasible !

Well, I'm not certain this'll help...

I haven't rendered the image yet.  I tried working with the CP editor to remove bad control points and then add new cps at what look like obvious matches.  Unfortunately, APG would not match overlapping push pins - sharp contrast, color distinct from surrounding area.  It also would not match sharp edges in this section of the subject.  Using the control point optimization tool only made it worse.  Notice the difference between RMS values in image 1 (post "optimization") and image 3 (pre optimization).  Image 3 is identical before and after CP editing.
The next step is to try just the subset of images that present the difficulties, again try to align obvious overlapping points, then render in both smartblend and multiband.  Another alternative is to just render what APG gives me, then cut and paste the good areas of both stitches together. :^) - I might try that, too.

Last edited by Apapane (2009-09-30 18:38:04)

hankkarl wrote:

Are you using a CPU or GPU?  If CPU, then in the pano editor window, check the settings.  you can select a blender there for the preview.

Hi Hank,

I have tried with and without GPU enabled - to no avail with the full 700 frames.  I took a 48 frame subset that included the badly stitched text.  After some CP editing, I tried rendering using Multiband and Smartblend with all 7 different interpolators.  Multiband bicubic gave OK results - no other interpolater worked well with Multiband rendering.
Smart Blend worked well with all interpolators except Nearest Neighbor.

I then went back to the full 700 frame GigaPan set, and cropped the projected image to the narrow band that included the poorly stitched text.  Even with more CP editing - no luck.

This vertical band that includes the text just happens to be the overlap area between the first and last columns of the captured frames.  This is where the GigaPan stitcher works very well.  Now, if there was a way to get the best of both worlds! :^)

Next step - cut and paste between GigaPan stitched pano and APG rendered pano.

Any other suggestions are welcome (except for "Don't take so many frames").
I continue "Beta" testing, even with a commercial product.

Richard

Georges,
Mahalo for your suggestions.

Here are more screen shots to show the situation with this pano in more detail.

The two identical pages are indeed separate, as shown in 1393-top-left, 1394-bottom-left, 713-top-right, and 714 - bottom-right.  This just happened to be the beginning and end of the frame capture (bad placement on my part).

The APG204-edit-mode series shows the entire 700 frames.  There were no links between 1393-713 and 1394-714 until I added control points in the CP editor.  These images would not line up correctly after more than 2 hours of trying.

So, I took a 48 frame subset of the original 700, and loaded, edited, and rendered those.  The text stitched just fine.  See the APG-subset series.

The "Subset48-post-find-cps-for-push-pins" shows that APG could not match what seem to be obvious control points with the yellow and green push pins, but the subset stitched just fine in spite of the lack of control points there.

See my first post in this thread to view screen shots that compare APG2.0.4 with the GigaPan stitched text & edges (from the entire 700 frames).  My conclusion so far:
Large panoramas with text at 360 junction (a rare occurrence, no doubt), GigaPan outperforms APG.
Most ANY panorama with many straight lines/edges, APG outperforms the GigaPan stitcher.  I look forward to improvements of both stitching programs.

OK - have at it!  I look forward to your insights.

Aloha,
Richard

Images:
1 - img1393-top-left
2 - img1394-bottom-left
3 - img713 - top-right
4 - img714 - bottom-right
5 - subset48-cps-close
6 - subset48-cps
7 - subset48-edit-mode
8 - litho-all-cps-close
9 - litho-all-cps-closest
10 - litho-all-cps
11 - litho-all-edit1
12- litho-all-edit2
13 - Subset48-post-find-cps-for-push-pins

Last edited by Apapane (2009-10-02 09:30:48)