dicengine
Could I choose ROI for stereo initialization？
My problem is a little like #120. Because part of the picture is not useful and some random pattern may cause large error such as glare. When I do some tests about cross-correlation. I found the cropped pictures have much better result if only considering cross-correlation. The Match in Paraview shows the cross-correlation result. But the problem is that when I crop the picture, it change the coordinate of each pixels and though I get good cross correlation, the model coordinate is wrong.
Cropped picture
Original picture
(There is a glare causing incorrect projection and it ruins the result.)
These are fm_projective_trans_0.png. I guess this is to calculate the 8 parameters thing we discussed last time. And right_projected_to_left_color.tif is the result of right plane projected to left plane with those 8 parameters.
So how can I draw ROI for this initilazition step? Is this possible? Or is it possible to use exhaustive search for cross-correlation? Because I struggle with cross correlation now.
In my understanding of cross-correlation prcedure. The initialization is really important in cross-correlation because if the initial guess is too far away from real value, the Newton-Raphson method will never jump there (jump_tolerance is only to avoid nonsense result instead of jump to find true value).
Am I correct? So if I want to improve cross-correlation, it is necessary to get a better initial guess. How close should the initial guess be away from real value?
You are correct about the initial guess, which is sometimes hard in stereo correlation. If you want a good cross-correlation, you need a good initial guess. The initial guess can be roughly 3-5pixels off and will often still converge. Having a better cross-correlation initialization method in DICe is something that is needed. For example if you have a non-planar surface, the cross-correlation will likely fail.
On Jun 25, 2019, at 11:48 AM, CaesarTheFox <[email protected]mailto:[email protected]> wrote:
In my understanding of cross-correlation prcedure. The initialization is really important in cross-correlation because if the initial guess is too far away from real value, the Newton-Raphson method will never jump there (jump_tolerance is only to avoid nonsense result instead of jump to find true value).
Am I correct? So if I want to improve cross-correlation, it is necessary to get a better initial guess. How close should the initial guess be away from real value?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHubhttps://github.com/dicengine/dice/issues/130?email_source=notifications&email_token=ADT7TORHMXH4FHKBFM2AF7TP4JK7FA5CNFSM4H3KBSAKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODYRBRZY#issuecomment-505551079, or mute the threadhttps://github.com/notifications/unsubscribe-auth/ADT7TOWMB2I6FLPUWUHORIDP4JK7FANCNFSM4H3KBSAA.
So could I choose ROI for stereo initialization？It influences the initialization significantly. Now I need to mask the part I dont want in Photoshop before processing. If we can have such a feature it would be more convenient I guess.