Train using our own data

[MajdWardeh]

Hi,

Thank you for this amazing paper!
I was wondering if you could provide the code to train the models.
I believe that the best way to prove the real performance of your approach is by letting people train it by themselves.

Thank you,

[RaresAmbrus]

Hi, thanks for the interest! Sadly we have no plans to release the training code, but happy to answer any specific questions about our method.

[MajdWardeh]

Hi @RaresAmbrus,

I am trying to implement the training part of KP3D. I have a couple of questions that I would be glad if you could help me with.

1. After estimating the rotation, R, by solving the Orthogonal Procrestous problem, the resulting R could have a det of -1. What do you do to resolve this? do you implement Kabsch algorithm?
2. For the geometric loss, It's stated that it is applied to the initial set of correspondences (the set obtained with descriptor matching). However, what about the outliers? is there any outlier rejection procedure or is there another rematching step (similar to KP2D)?
3. For the descriptor loss, What are the positive descriptors associated with the anchor descriptors set? Are they defined by the inliers set coming from solving PnP with RANSAC? Otherwise, how could we find those positive descriptors?

Thank you so much (and for your colleagues) for this great paper!
Kind regards,
Majd Wardeh

[MajdWardeh]

Hello @RaresAmbrus,

I am sure you are very busy. I would be glad if you could help me clarify the questions above when you have the time.

Kind regards,
Majd

[Cc19245]

Hello @RaresAmbrus,

I am sure you are very busy. I would be glad if you could help me clarify the questions above when you have the time.

Kind regards, Majd

Hi, Majd! I'm also struggle to reproduce the training code of KP3D. Maybe we can get some communication? My email is [email protected]. Looking forward to you reply!

[RaresAmbrus]

Hi, thanks for the interest and sorry about the slow reply, I'm out of the office and will only get back around the end of Nov. I'll try to take a look at the code. In the meantime, maybe @jiexiong2016 is available?