gmmtree_rigid.py example failed to register about probreg

[neka-nat]

I recognize that accuracy of gmmtree is not good in the implemented algorithms.
So I'm debugging.
However, the gif animation attached to README is the result of actually executing gmmtree_rigid.py.
Is your result worse than this?

[YungKC]

Hi neka-nat, attached is the screenshot of the result that I got: Btw, thank you for putting together this github repository. We are researching various ways to register noisy 3D point clouds quickly and accurately. One area we are looking into is HGMM described by Eckart et al from Nvidia where the latest version claimed to perform each E step in less than 10 ms with their clever use of indexing to fit within GPU's shared memory. However, we are taken back by the local minima issues. The Support Vector Registration algorithm looks more promising in that regard. Are you interested in helping us in the our feasibility study and prototypes? We are a small startup with goals to achieve a real time 3D pose tracking of well defined objects. We can make accommodations and can compensate for your effort. Look forward in hearing back from you. Kai Yung EBots Inc.

…

On Thu, May 16, 2019 at 6:07 AM neka-nat ***@***.***> wrote: I recognize that accuracy of gmmtree is not good in the implemented algorithm. So I'm debugging. However, the gif animation attached to README <https://github.com/neka-nat/probreg/blob/master/images/gmmtree_rigid.gif> is the result of actually executing gmmtree_rigid.py. Is your result worse than this? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AADWZ5C6B7U57SKTN7DHCTDPVVMCRA5CNFSM4HNGFHDKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODVRX7SY#issuecomment-493060043>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AADWZ5EQ5UZXB4NTVW3MTITPVVMCRANCNFSM4HNGFHDA> .

-- Kai

[neka-nat]

That's great!
I'm very interested in the implementation of HGMM on a GPU.
SVR is weak for simple Gaussian noise, but I have the impression that it is strong for real problems.

I'm also interested in helping you, but I can only help in a little time and remotely.
If I can help you in that way, I hope to do.

[YungKC]

Hi, Yes we can accommodate your remote location and your schedule. I would like to get an idea how much time can you carve out to work on this project. Our goal is to reach real time performance on 3d pose determination from a reasonably sized point cloud, and I can describe the problem in more detail once we work out the terms of the project. I also discovered filterReg from your GitHub project that I was not aware of before. Very promising. Can you suggest a 20 minute window in which we can discuss more about this? I am also interested in knowing your background and what type of arrangements you are comfortable with. Look forward to hearing from you. Kai Yung Ebots

…

On Sat, May 18, 2019, 8:21 AM neka-nat ***@***.***> wrote: That's great! I'm very interested in the implementation of HGMM on a GPU. SVR is weak for simple Gaussian noise, but I have the impression that it is strong for real problems. I'm also interested in helping you, but I can only help in a little time and remotely. If I can help you in that way, I hope to do. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#1?email_source=notifications&email_token=AADWZ5AJWFYVYRHHB7MAXGDPWANGBA5CNFSM4HNGFHDKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODVWQQ5Q#issuecomment-493684854>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AADWZ5AELLWAO357W534U6DPWANGBANCNFSM4HNGFHDA> .

[neka-nat]

Sounds good.
Could you send it to my email?
[email protected]
I will send you my background material in return and discuss the schedule of the teleconference.

[neka-nat]

The transformation using twist has been changed to the method using the Rodrigues rotation formula, not by linear approximation.
This has improved the accuracy.