Forum

Author Topic: Keypoints-Tie points, calibration parameters and optimize..tests and doubts  (Read 6105 times)

Outis79

  • Newbie
  • *
  • Posts: 18
    • View Profile
Hello everybody,
I know that some posts have already been written on such issues but I've still some doubts.
I'm currently working on an aerial photogrammetry project and I've tried some tests.

1st test, Camera calibration parameters.
The camera (ILCE-7 35mm fixed lens) has been calibrated into Agisoft Lens. I've performed the alignment of a subset of photos 3 times: precalibrated with "fix calibration" on, precalibrated with "fix calibration" off, auto.
After the alignment (high accuracy, generic preselection) with the last two modes photoscan gives me into the "Adjust tab" new intrinsic parameters which are pretty different from the Pre-calibrated ones. As said in other discussions on the foum, due to the different environment conditions and difference of focus settings between the calibration process with Agisoft Lens and the shooting phase we can get different values. Moreover looking at the three chunks "show info", the first mode gives higher RMS and Max reprojection errors while the second and the third mode give very similar errors.   
There's a difference between the 2nd and the 3rd mode, the latter doesn't calculate any k3 and k4 parameter.
So the adjusted parameters are more reliable and that we need pre-calibrated parameters only to make photoscan consider the k3 and k4 parameters, can you confirm that?
Since the k3 and k4 parameters are not 0, must they be set on "fixed" into the next optimize process?

2nd test. Alignment KeyPoints and Tie Points number.
I've lunched for the same images subset 4 different alignments varying the number of Key Pt and Tie Pt: 40k-4k, 40k-0, 80k-0, 120k-0.
The number of tie points I got is between 120k of the 1st chunk to over 2 milions of the last chunk. I can notice by the Show_info that the more the KeyPt-TiePt number the lower the Reproj Error. The difference is not very high (0.9 in the 1st, 0.7 in the 3rd and 4th) but I've noticed that using the gradual selection and putting the limit of 10-15% of TiePt that can be deleted, I can reach the 0.5px of RMS Reproj error in the 2nd and 3rd chunk.
It might imply that using a higher Key Points number and 0 Tie points give you the chance to delete a lower percentage of points and to reach a lower reprojection error by the gradual selection and optimizing. What do you think?
Using the gradual selection Image Count bar I would like to remove all the Tie pts which are only in 2 images since I think they are responsible of much of the noise. In all the chunks I would remove almost 50% of all points which is maybe a too much high percentage but into the 4th chunk (120k-0) I would still keep over 1 million of points that is far higher than the total number of pts I have in the 1st and 2nd alignment. Can I remove so many points after setting a higher number of Key Pts and Tie Pts in the previous alignment process?

If I'm going to perform an optimize based on GCP what's the need of the gradual selection optimize....I mean does the GCP optimize delete the previous gradual selection optimize or Photoscan gives them some type of weights?

Regarding the "update tool" after placing the GCPs, what kind of transformation (affine,conform...) can be performed?

Thank you all.     

Outis79

  • Newbie
  • *
  • Posts: 18
    • View Profile
Nobody? :(

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14927
    • View Profile
Hello outis79,

If you are running the Align Photos operation with the disabled option for Adaptive Camera Model Fitting, PhotoScan will adjust only certain set of parameters: f, cx, cy, p1, p2, k1 - k3. However, if some precalibrated values are different from zero, they will be adjusted as well.

Optimization process doesn't take into account the removed tie points, so usually removing point outliers and points with high reprojecton error is recommended prior to the optimization process.

As for Update tool, it uses only Affine transformations (translation, rotation and scale), thus keeping relative positions of the cameras unchanged.
Best regards,
Alexey Pasumansky,
Agisoft LLC

Outis79

  • Newbie
  • *
  • Posts: 18
    • View Profile
Thanks Alexey,
your reply answers some questions. Athough I wonder also weather is good to remove points wich are projected by only 2 photos using the gradual selection and, as far as I noticed, are responsible of noise.
It has been already said in other discussions to avoid to remove too many tie points (10% some,1/5 others). Usually I don't remove more than 10%-15%, but I wonder if I can remove up to 50% in such a case:
1) I have a large sparse point cloud obtained by a higher number of keypoints and 0 limit of tie points into the alignement process
2) the tie points to get cancelled are well spread over all the project area
3) I keep a large number of tie points and checking they are well distributed in all photos. I might use the "Photo>View Points" to do that.
Thanks