Your post is interesting, but you do not provide enough information for it to be usefull.
Could you please provide
. your environment details (Operating system,...)
. one pano made with 1.4.2 , together with the main detection and optimise parameters
. the same as above with APG 2.0.1
Bernard

very nice Bernard,

I can see your frustration and that is really weird it is happening.

thanks for sharing

Henrik

bernardlanguillier wrote:

The lens was rotated around its foot collar using a Winberley side Kick. It is therefore not rotated around the entrance pupil, but the main subject is probably at least 200m away from the camera.

There is no main subject for Autopano, even if they are fuzzy or suffer from parallax it will attempt to stitch the nearby parts of the subject as well as the distant parts...

It's very difficult to find out the actual source of your problems without knowing the RMS of the links resulting from using the old and the new versions of Autopano. I suspect some links in the leftmost part having a rather bad RMS. This can disturb the whole pano. If you want to verify my suppositions you could try to stitch the source images corresponding to the center part of your pano only. The clouds in the rightmost part could cause some problems, too.

Gurl what you say is correct but APP 1.4.2 has been so perfect even with hand held images V2 is not so forgiving.

He shot a distant scene that too with 300 f2.8 that should work even if I shoot hand held, at lesat 1.4.2 worked flawlessly in such cases.

Thanks for sharing this example, quite useful !
I must say I am quite happy with 1.4.2, and still hesitant to move to V2.

I am using smartblend most of the time, what is exactly  the advantage of using multiband when there are no 'moving' objects ?
Bernard

GURL wrote:

bernardlanguillier wrote:

The lens was rotated around its foot collar using a Winberley side Kick. It is therefore not rotated around the entrance pupil, but the main subject is probably at least 200m away from the camera.

There is no main subject for Autopano, even if they are fuzzy or suffer from parallax it will attempt to stitch the nearby parts of the subject as well as the distant parts...

It's very difficult to find out the actual source of your problems without knowing the RMS of the links resulting from using the old and the new versions of Autopano. I suspect some links in the leftmost part having a rather bad RMS. This can disturb the whole pano. If you want to verify my suppositions you could try to stitch the source images corresponding to the center part of your pano only. The clouds in the rightmost part could cause some problems, too.

Maybe, but the RMS is 2.12, which is excellent and there was no bad link. The proof of this is clearly that there is no problem with multi-band for this image. Even the close end of the scene was probably 50m away from me anyway and PTgui didn't break a sweat stitching this with an optimization result of 0.5 after the first optimization without any point deletion (which is in fact remarkably good and a very rare occurence even when the nodal point is used within 0.5 mm).

Besides, the point is that there was a clear regression going from 1.42 to 2.02 about the most important aspect of the application, which is stitching quality. I couldn't care less about GPU acceleration and faster performance if the stitch is useless.

Multi-band is a valid option for such images, but delivers really poor results in clouds with a lot more blurring compared to Smartblend.

The main quality of 1.42, stitch an forget ability, is gone and that is a MAJOR issue. One that will personnally make me stop using the application until it is recognized fix.

cheers,
Bernard

Last edited by bernardlanguillier (2009-06-16 09:10:10)

I'm sure it's just a matter of setting which have changed between version. I measured the quality in RMS and visually over our large database and most of time, the new version do better ( in RMS quality and visually ). Don't forget my note about RMS on fisheye which is a different problem.

So let's just found the setting that leads a some misaligned in v2.0.2 ( even if this zone in near pixel wide ).
Could it be possible to make public the posted sample sets to play with ?

http://www.autopano.net/forum/t766-ftp-server

OK, will do tonight. I guess I create a folder with my name or something like that?

In fact, folder creation on the ftp is not allowed because it would say that listing is allowed ( or we want to keep your file secret to any other user ). That's why you need to zip your package and upload a zip ( in the zip, you can store the folder structure ). BTW : you can use zip but also tar.gz, gzip, bzip2.

Sorry for the delay in the answer.
Here's the result of my analysis. It's just the same quality.

How I did my tests :
- 1.4.2 : default setting with lens distortion optimization checked on
- 2.0.2 : optimization settings preset set to strong. Everything else in default mode

Render under PSB with layer to visualize quality of layer overlapping.

Screenshot 1 : v1.4.2 with the layer we want to analyse on over the smartblend result.
So we can see a slight difference between both images which is normal ( we are at 60 %, 70 % of the 100 % size ).

Screenshot 2 : v1.4.2 smartblend panorama result only. We cannot see any issue in stitching.

Screenshot 3 : v2.0.2 with the layer we want to analyse on over the smartblend result.
Quite exactly the same result as with 1.4.2. Pixels have not more difference than in the 1.4.2

Screenshot 2 : v2.0.2 smartblend panorama result only. We cannot see any issue in stitching.

I didn't change that part between both version and that's what I was first thinking. I was thinking so also, but found out that calculated distortion were the same.