[APG 2.5.0 A1] Poor FE detection and unstable/diverging optimization  

Archive of all bug reports. This forum is closed - you cannot create new topics or comments here! If you think a specific topic was moved here by mistake, please contact the moderators!
no avatar
enbilaman
Member
 
Posts: 91
Joined: Thu May 08, 2008 6:50 pm

[APG 2.5.0 A1] Poor FE detection and unstable/diverging optimization

by enbilaman » Sat Jun 26, 2010 8:09 pm

1) Test with simple panoramas. All are composed of four fisheye images with Canon FF + Tokina 10-17 @ 12 mm on one row. MacOS 10.6 32-bits.

[APG 2.0.9] gives a "perfect" result just by pushing on the detection button.
Then, without any change, making a further re-optimization by pushing the button on the CP Editor tab window either improves the RMS or it lets it at the same value and the panorama is kept subsequently in the same perfect appearance whatever the number of additional re-optimizations. The process seems to converge toward the best or optimal outcome.
The result is comparable with different sets of images. Excellent output.

[APG 2.5.0 A1] gives a poorer result compared to the previous test with 2.0.9, ranging from "perfect" to "very good" depending on the set of images and on something else (the RMS result varies slightly but randomly while using the same set of image).
Then, without any change, making a further re-optimization (by pushing the button on the CP Editor tab window) always induces a regression. One additional re-optimization degrades sensibly the panorama. After two re-optimizations, more than often an image has become orphan! After more re-optimization, all the images are deprived of all their detected CPs!

2) Test with HR panoramas. All are composed of two rows =/- 45° plus Nadir and Zenith (i.e. 14 images). Same combo as before but the lens is zoomed at @ 17 mm focal length.

[APG 2.0.9] gives an "perfect" or "very good" result depending mainly on the set of source images; "17 mm/Fisheye" selection (but no FE circle adjustment) had been done before pushing on the detection button.

[APG 2.5.0 A1] gives a much more mediocre RMS result compared to the previous test with 2.0.9, ranging from "good" to "not good" (or worst) depending on the panorama and on something else: the result is never at all the same while inputing the same set of images and same process though: RMS varies randomly from min.=5 to about 20 in a series of ten successive tries!
Then, without any change, making a further re-optimization by pushing the button on the CP Editor tab window always induces a rapid and sharp divergence: one additional re-optimization degrades dramatically the panorama and some images may loose their links. The panorama is broken into several sub panoramas all of them being quite messy. After one more re-optimization... it's an indescribable Contemporary Art master piece!

Michel

no avatar
digipano
Member
 
Posts: 581
Joined: Sat Feb 16, 2008 9:07 am

by digipano » Sun Jun 27, 2010 4:26 am

Yes I also observed the same [APG 2.5.0 A1] has not been able to do any of my test panorama from simple to complex ones all throw up garbage & have to manually intervene.

The strange part is the detection finds control points to non related images on non related objects which are no way similar.

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Sun Jun 27, 2010 9:42 am

In this example 3 clicks in a row on the Detection button resulted in 3 different RMS: 12.8, 16.9 and 19.4...


Last edited by GURL on Sun Jun 27, 2010 9:43 am, edited 1 time in total.
Georges

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Sun Jun 27, 2010 9:52 am

A new series of attempts - I closed and reopened APG - resulted in 3 different values: 14.2, 11.4 (THE BEST ONE) and 15.0...


Georges

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Sun Jun 27, 2010 10:10 am

The strange default settings values Autopano is using for fisheyes only is in my opinion enough to explain why several optimizations in a row (not several detections) are causing RMS changes: it delete links where RMS is higher than 10 but don't optimize again after deleting these links.


Last edited by GURL on Sun Jun 27, 2010 10:56 am, edited 1 time in total.
Georges

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Sun Jun 27, 2010 10:27 am

Rotating the pano so that the nadir is in the middle of image shows that a link between every image pair was created ( ;) the way I use to delete them easily...)

The worst link shows very unusual CP locations.

(This pano is yet an impossible example of automatic stitch for many reasons - I used APG on Win 7 / 64 bit.)




Last edited by GURL on Sun Jun 27, 2010 10:33 am, edited 1 time in total.
Georges

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Sun Jun 27, 2010 11:29 am

Here is a possible WORKAROUND to stitch this pano.

It's non automatic, it includes 9 steps and I think this is too much but you can see (well, you must believe me) that I didn't take care of individual control points (even those stupid ones in the sky) but only deleted 17 useless links (even a green one) keeping only 8 useful ones, optimized, deleted the worst CP from all images, optimized, deleted the worst CP from all images, optimized, etc... (if you delete too much CP just use History.)
















Last edited by GURL on Sun Jun 27, 2010 12:11 pm, edited 1 time in total.
Georges

no avatar
enbilaman
Member
 
Posts: 91
Joined: Thu May 08, 2008 6:50 pm

by enbilaman » Mon Jun 28, 2010 9:34 am

GURL wrote:The strange default settings values Autopano is using for fisheyes only is in my opinion enough to explain why several optimizations in a row (not several detections) are causing RMS changes: it delete links where RMS is higher than 10 but don't optimize again after deleting these links.

Georges,

Of course, I had also thought that weird default values could cause the problem. But I confirm that both active "Advanced Options" had been (apparently) set exactly the same on both APG 2.0.9 and 2.5.0 A1 after having pushed the reset-to-default button. I have used these default settings from then on. So I believe that the root cause of the flaw is elsewhere unless there is an anomaly in the "setting options" software module of Alpha1.

Michel

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Mon Jun 28, 2010 11:18 am

enbilaman wrote:Of course, I had also thought that weird default values could cause the problem. But I confirm that both active "Advanced Options" had been (apparently) set exactly the same on both APG 2.0.9 and 2.5.0 A1 after having pushed the reset-to-default button. I have used these default settings from then on. So I believe that the root cause of the flaw is elsewhere unless there is an anomaly in the "setting options" software module of Alpha1.

Well, I think there is something very unclear...

When Autopano knows that the lens is a fisheye, rather than selecting the "standard" default parameters it uses the one shown in the first capture (and they are strange, IMHO)

Then, if you push the Restore Defaults button you get ordinary (non fisheye) default values- shown in the second capture - and this is a dubious behavior as the lens is still considered as being a fisheye...

I believe there should be several default values (for example one for fisheye, one to be used when many source images have been loaded (mosaic) and a third for ordinary panos.)

Though Autopano is not supposed to be used this way, for 360x180 panos I always look first at automatic circular crop circles and alway adjust optimization settings before hitting the Detect button because by doing so I avoid many failed attempts!




Georges

User avatar
AlexandreJ
Kolor Team
 
Posts: 5918
Joined: Mon Nov 14, 2005 4:56 pm
Location: Francin, France

by AlexandreJ » Mon Jun 28, 2010 2:38 pm

Issue 125 opened

no avatar
klausesser
Member
 
Posts: 7886
Joined: Mon May 22, 2006 12:18 am
Location: Duesseldorf, Germany

by klausesser » Mon Jun 28, 2010 2:43 pm

AlexandreJ wrote:Issue 125 opened

:cool:

Btw.: is "lens distortion" meant for fisheyes only (i read something like that before)?
and "advanced lens distortion" meant for off-centered fisheyes?

I have a feeling that my 20mm Nikon on a fullframe is mostly misinterpreted - never get a sufficient stitch without
fighting for hours with CPs correction . . . and some more time retouching remaining errors afterwards . . .

tried the manual CP correction in 2.5 - but it crashes all the time (MacBook Pro, 2x2.8GHz, 8GB, SnowLeopard 10.6.4)

best, Klaus
Last edited by klausesser on Mon Jun 28, 2010 2:49 pm, edited 1 time in total.
Simplicity is the keynote of all true elegance. Coco Chanel

User avatar
[bo]
Member
 
Posts: 1226
Joined: Fri May 05, 2006 8:16 am
Location: Bulgaria

by [bo] » Mon Jun 28, 2010 2:44 pm

It never seizes to amaze me how GURL can transform a simple thread into some kind of scientific research publication :D Just thinking of the time it takes to take annotate and post screen shots, let alone test many different scenarios - it makes me wonder if he has some control over time-space continuum?
Some of my panoramas, posted in the Autopano Pro flickr group.

no avatar
GURL
Member
 
Posts: 2943
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble

by GURL » Mon Jun 28, 2010 4:20 pm

'[bo wrote:']It never seizes to amaze me how GURL [etc]

:lol: When seen as a computer game stitching is really fun (much more than cross-words or such hobbies.)

BTW, I posted only some of my attempts with Piazza San Pietro and APP 2.5 as I can't understand yet why the best one are so good ! :cool:
Georges


Return to Archive

Who is online

Users browsing this forum: No registered users and 1 guest