search

zhulf0804 / GCNet

Leveraging Inlier Correspondences Proportion for Point Cloud Registration. https://arxiv.org/abs/2201.12094.
MIT License
96 stars 12 forks source link

Perfomance #9

Closed XMU-qcj closed 2 years ago

XMU-qcj commented 2 years ago

Hello, thank you for your work, I encountered a problem, I used the pre-trained model and eval code provided by you to test, and I could not restore the effect provided by your paper on the MVP_RG data set. Where are the details that need to be changed for testing? ?

zhulf0804 commented 2 years ago

Hello, thanks for your interest in this work.

There are some randomness in RANSAC operation. So we can get a simiar result when testing. The result on the MVP_RG reported in the paper is easy to reproduce, because we didn't report the best result which occurs occasionally.

We used Open3d v0.10 to test. We found different performance were achived with different versions of Open3d, because RANSAC implemtations are different in differetn Open3d versions. What is your Open3D version?

XMU-qcj commented 2 years ago

Hello, thanks for your interest in this work.

There are some randomness in RANSAC operation. So we can get a simiar result when testing. The result on the MVP_RG reported in the paper is easy to reproduce, because we didn't report the best result which occurs occasionally.

We used Open3d v0.10 to test. We found different performance were achived with different versions of Open3d, because RANSAC implemtations are different in differetn Open3d versions. What is your Open3D version?

Yes, I realized the randomness, I am using the same version of open3d as you, and I have also done multiple tests, but still can't get good results, most of the rotation errors stay above 30. And I noticed that the number of npts set in your eval code is 768, will this affect the effect?

zhulf0804 commented 2 years ago

Hello,

I tested twice just now with the README guidance (without modifying code), and the results (RRE 7.74 and 8.47) are as follows:

mvp_rg_res

Did you test with the similar command ?

XMU-qcj commented 2 years ago

Hello,

  • rotation errors stay above 30 is not reasonable, maybe there are improper operations.
  • the number of npts in the code affects, but it doesn't affect too much.

I tested twice just now with the README guidance (without modifying code), and the results (RRE 7.74 and 8.47) are as follows:

mvp_rg_res

Did you test with the similar command ?

I changed a server, used 3090 instead of 1080, and performed the same environment configuration and test, this time the result is excellent, which is very strange, but thank you for your reply

zhulf0804 commented 2 years ago

Yes, it's strange.

We have tested the performance using different machines and GPUs, and achieved a similar performance.