Forum

Author Topic: Tie points once again  (Read 7698 times)

janoz

  • Newbie
  • *
  • Posts: 8
    • View Profile
Tie points once again
« on: March 05, 2015, 01:12:34 AM »
Having used Tie points config option in v.1.1 for some while I am stil having some problems with understanding this feature. In many cases leaving the default setting at 1000 causes too sparse "sparse point cloud" leading in next step to some inaccuracies in generating dense point cloud espacially in too thin parts of the sparse point cloud. These problems are  mostly solvable by disabling tie point limit feature (setting limit at 0). Is there any additional benefit by the implementation of tie point limit beside speeding up the process (I think in one of the posts it is stated that tie point config automatically filters out "badly" projected points but is  this objective not achievable in better or more controlled way by the application of traditional reprojection error decreasment?) . Is there any other desirable tie point limit setting beside 1000 that would turn out as a good presupposition for dense point cloud generation.
And what is exactly the relation of tie points to key points. If sparse point cloud is determined solely by tie point config why to bother any more about key point config?
I am probably missing some  substantial knowledge on this topic but explanation in user manual especially on tie point is too theoretical for me :(

aletopo

  • Guest
Re: Tie points once again
« Reply #1 on: March 06, 2015, 03:42:56 PM »
I already wondered about the same questions. I would have expected your sparse point cloud being even thinner by disabling tie point limit, because then you can get e.g. an image pair with 5000 tie points (instead of 1000) which means that only image pairs with more than 10% = 500 tie points (instead of 100 tie points) are considered for dense cloud model. That is what i understand from the manual.
In my suggestion there is no benefit in higher limit because 1000 tie points are more than enough for accurate alignment.
Concerning key points I think Agisoft finds characteristic points (=key points) in each photo individually and then tries to match them with key points found in other images. If a match is succesful you get a tie point. Please somebody correct me if I am wrong here.

Marcel

  • Sr. Member
  • ****
  • Posts: 309
    • View Profile
Re: Tie points once again
« Reply #2 on: March 07, 2015, 12:38:49 AM »
I think in one of the posts it is stated that tie point config automatically filters out "badly" projected points

That is actually an interesting idea. Perhaps Alexey could confirm this? When the Tie point setting is active, does Photoscan rate the quality of the points and use only the highest quality points?

Personally I set it to 0. The alignment never takes a very long time anyway  (compared to the Dense Cloud reconstruction). I rather have it run for an hour and be sure that it has the best alignment possible.

JMR

  • Hero Member
  • *****
  • Posts: 504
    • View Profile
Re: Tie points once again
« Reply #3 on: March 09, 2015, 03:05:09 PM »
I personally find that the default 10.000 limit causes bad alignment in too much cases. I wonder how the quality assessment is done.
If reprojection error is the only factor... What if the initial guess of camera orientations (internal and external) is not good enough?
In my experience, the Initial guess (before optimizing) almost always needs refinement. If that is the case for most consumer (non metric) cameras, it is likely to happen that tie-points filtering before optimization may lead to kill a lot of them for no good reason, specially near the frame borders where complex distortions are bigger. This seems to me to lead to alignment failures.
Can Alexey give us a good clue to understand the need for thinning the sparse cloud? thanks in advance.

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14927
    • View Profile
Re: Tie points once again
« Reply #4 on: March 17, 2015, 05:40:24 PM »
Hello,

Thin point cloud tool is trying to remove first of all tie-points with the lower number of projections. Reprojection error is not used here, since tie-point limit on alignment stage is applied before the camera alignment is estimated.


Using tie-point limit is reasonable for extremely big projects with thousands of images, as it will influence on alignment time and also on the response time in the Model view in sparse cloud view mode.
« Last Edit: March 17, 2015, 05:41:58 PM by Alexey Pasumansky »
Best regards,
Alexey Pasumansky,
Agisoft LLC