Impact of number of Control Points and Ransac model?  

This forum is dedicated to Autopano Pro / Giga bug reports and features discussion.
Please read the posting rules before starting a topic!
no avatar
pstein
Member
 
Topic author
Posts: 38
Likes: 0 post
Liked in: 0 post
Joined: Thu May 13, 2010 1:17 pm
Info

Impact of number of Control Points and Ransac model?

by pstein » Thu Jul 07, 2011 5:44 am

In Preferences/Detection tab there is an option to define the number of Control points.

From my point of view 10 Control Points are enough.
AP should have with 10 CP all information to paste them successfully together.

Choosing a value of e.g. 200 is overkill.

Am I right?

Is decrease of processing speed the only dis-advantage when choosing a high value?

Furthermore I would like to know your experience of the two "Ransac models" here: "Similarity" vs. "Homography"
Whats the difference?

When should the user take which Ransac model?

Is everything trial and error? No rules of thumb?

Peter

User avatar
AlexandreJ
Kolor Team
 
Posts: 5987
Likes: 7 posts
Liked in: 10 posts
Joined: Mon Nov 14, 2005 4:56 pm
Location: Francin, France
Info

by AlexandreJ » Thu Jul 07, 2011 8:55 am

pstein wrote:From my point of view 10 Control Points are enough.
AP should have with 10 CP all information to paste them successfully together.

It depends and that was true for example with PtGui.
The design or Autopano is different and it's the power of the cp detector with a massive amount of CP ( and then statistical results ) that gives the quality of stitch.

The more CP you have, the more you will be able to find moving object, ghost, everything that moves and should not. On a static scene, 10 CPs should be enough but you need to be sure that all the CPs are really well located.

pstein wrote:Choosing a value of e.g. 200 is overkill.

No. It takes more time, can help you for some complicated case, it's just about that.
Note : We are in fact generating between 50 et 2000 Cps for each image pair in autopano. We decimate them to prevent using really all of them ( too much ).

pstein wrote:Is decrease of processing speed the only dis-advantage when choosing a high value?

The speed is quite linear with the number of cps. But because it's well optimized, you should not see the effect of raising the CPs number much.
You will notice a far bigger impact by changing the detection quality.

pstein wrote:Furthermore I would like to know your experience of the two "Ransac models" here: "Similarity" vs. "Homography" Whats the difference?

* Similarity is for focal length above 40, 50mm. It's fast, has a really good detection quality and does perfect result for such lenses.
* Homography is needed for short focal length or do multiple viewpoint detection. It's a bit slower and doesn't have the same detection quality as similarity. We are still improving this algorithm to get the same quality as Similarity, but we are not there yet.

no avatar
Photomaker
Member
 
Posts: 10
Likes: 0 post
Liked in: 0 post
Joined: Wed Jan 26, 2011 11:17 pm
Info

by Photomaker » Fri Jul 08, 2011 10:29 pm

Hello,

I have two questions concerning the control points and/or the detection quality.

I tried to stitch 3 photos and the CP editor show me for 2 pictures the attached result:

Most of the CPs are present in a very small range of the overlapping zone of both pictures. Is there a possibility to force a broader zone, where CPs are located?

A few CPs are located in the sky. Is it possibel to remove such CPs automatically? When I press the button "Delete bad CPs" the red point in the sky still remains.

Best regards

Photomaker



no avatar
Photomaker
Member
 
Posts: 10
Likes: 0 post
Liked in: 0 post
Joined: Wed Jan 26, 2011 11:17 pm
Info

by Photomaker » Fri Jul 08, 2011 10:45 pm

There is a second issue, I want to ask:

When I change the detection quality for recognizing photos that belong to a pano, there is the possibility to choose "low", "standard" and "high".

I tried all three detection qualities with a set of 6 photos. The resulting panos (see attached photos) show different RMS, i.e. standard: 2,87, low: 3,68, high: 4,35. Hence I would expect that the pano generated with a high detection quality is the poorest one, but interestingly it is the only one which do not show problems in the stitching (see upper part of attached photos). The first photo is the photo with the low detection quality, the last that with the high detection quality. Interestingly, in the pano with standard detection quality, the persons in the center of the pano disappeared.
In another series, the high detection quality results in poorer stitching quality but lower RMS as compared to the standard detection quality.
I used always 200 CPs for the detection. Is there a general advice possible, what is a good combination of number of CPs and detection quality?







no avatar
mediavets
Moderator
 
Posts: 16415
Likes: 2 posts
Liked in: 130 posts
Joined: Wed Nov 14, 2007 2:12 pm
Location: Isleham, Cambridgeshire, UK.
Info

by mediavets » Fri Jul 08, 2011 10:54 pm

Photomaker wrote:I used always 200 CPs for the detection. Is there a general advice possible, what is a good combination of number of CPs and detection quality?

Unless you have some reason to try something different stick with the install defaults.
Andrew Stephens
Many different Nodal Ninja and Agnos pano heads. Merlin/Panogear mount with Papywizard on Nokia Internet tablets.
Nikon D5100 and D40, Sigma 8mm f3.5 FE, Nikon 10.5mm FE, 35mm, 50mm, 18-55mm, 70-210mm. Promote control.

no avatar
GURL
Member
 
Posts: 2943
Likes: 0 post
Liked in: 0 post
Joined: Tue Dec 06, 2005 1:57 pm
Location: Grenoble
Info

by GURL » Sat Jul 09, 2011 6:24 pm

mediavets wrote:
Photomaker wrote:I used always 200 CPs for the detection. Is there a general advice possible, what is a good combination of number of CPs and detection quality?

Unless you have some reason to try something different stick with the install defaults.

Image

:) Non, la curiosité n'est pas un vilain défaut !

Pstein or Photomaker's questions and AlexandreJ's answers are interesting ones...
Last edited by GURL on Sat Jul 09, 2011 6:30 pm, edited 1 time in total.
Georges

User avatar
AlexandreJ
Kolor Team
 
Posts: 5987
Likes: 7 posts
Liked in: 10 posts
Joined: Mon Nov 14, 2005 4:56 pm
Location: Francin, France
Info

by AlexandreJ » Mon Jul 11, 2011 10:13 am

Photomaker wrote:Most of the CPs are present in a very small range of the overlapping zone of both pictures. Is there a possibility to force a broader zone, where CPs are located?

It depends on the Ransac model and on the images.
For long focal with default Ransac model ( ie simularity ), you'll get a really large overlapping.
For short focal with default Ransac model, you'll get this kind of result. If you want to have CP over all overlapping zone, you need 'homography'.

So why didn't we put homography by default ? Because it's slower and in many case doesn't have the same detection level as similarity ( which detects situation that homography doesn't ).
We are still working on this part to improve in all case homography. We will also propose something that will select automatically the better by using focal length as a decision criterion.

Photomaker wrote:A few CPs are located in the sky. Is it possibel to remove such CPs automatically? When I press the button "Delete bad CPs" the red point in the sky still remains.

Don't care about them too much. Our algorithms are robust to wrong CP. It doesn't influence at all the end quality.

User avatar
AlexandreJ
Kolor Team
 
Posts: 5987
Likes: 7 posts
Liked in: 10 posts
Joined: Mon Nov 14, 2005 4:56 pm
Location: Francin, France
Info

by AlexandreJ » Mon Jul 11, 2011 10:19 am

Photomaker wrote:When I change the detection quality for recognizing photos that belong to a pano, there is the possibility to choose "low", "standard" and "high".

I tried all three detection qualities with a set of 6 photos. The resulting panos (see attached photos) show different RMS, i.e. standard: 2,87, low: 3,68, high: 4,35. Hence I would expect that the pano generated with a high detection quality is the poorest one, but interestingly it is the only one which do not show problems in the stitching (see upper part of attached photos). The first photo is the photo with the low detection quality, the last that with the high detection quality. Interestingly, in the pano with standard detection quality, the persons in the center of the pano disappeared.
In another series, the high detection quality results in poorer stitching quality but lower RMS as compared to the standard detection quality.
I used always 200 CPs for the detection. Is there a general advice possible, what is a good combination of number of CPs and detection quality?

Here's the quality of detection and number of CPs is given different end RMS.

Quality of detection is mostly about the fact that overlapping percentage is low or high. If some panorama has too low overlapping, you'll need high detection quality. It won't really raise the panorama quality, but detection quality which will look further for overlapping.

CP number : if you raise this value without coping with other lens distorsion factor like offset in center of the lens, second and 3rd order distorsion, you'll end up with a RMS a bit higher because you have more CP and no option to model all these. With the new upcoming 2.6 version, raising this value and given new freedom to lens distortion leads to a lower RMS. And that's a really good news because it is what you've expect.


Who is online

Users browsing this forum: No registered users and 5 guests

cron