Results for rotated 3DMatch/ 3DLoMatch dataset

[wendyyuu]

Thank you for the insightful explanation you shared earlier. I have a few questions related to points 1 and 2 that I would like to clarify.

Just to confirm, are you suggesting that the random rotation, which is generated using the method described in point 1, is applied exclusively to the ground truth point pairs in various algorithms, like Predator?

Additionally, I'm curious about the default setting in YOHO. Is the default point cloud used for YOHO's training and evaluation the original 3Dmatch and 3DLomatch, or is it the rotated version? Upon examining the code, it seems that the random_rotation_matrix() function is only called during the training phase.

I appreciate your assistance in addressing these inquiries.

(1) The random rotation we add comes from the random rotation vector(utils.utils-->random_rotation_matrix()).
(2) Instead of recomputing the info file, the computational process of RR for the rotated version is performed directly by selecting the ground truth point pairs.​
(3) It could be the case that the FMR and IR results are similar in both dataset versions, but the RR has decreased (using the .info file recomputed), I think it should be the approximate calculation that exists in the calculation process of the info file, leading to an error in the case of random rotation.
(4) Another simple but robust way to verify is to compare the ground truth transformation and the predicted transformation directly, just like pointDSC, without considering the directly adjacent point cloud pairs, the results could be similar.

Originally posted by @HpWang-whu in #2 (comment)

[HpWang-whu]

Hi @wendyyuu,
Thanks for your interest!

1. We add random rotations to each point cloud. Then we conduct feature extraction and registration on the selected scan pairs. To calculate RR on rotated 3D(Lo)Match, the ground truth point pairs are selected on rotated scans.
2. The 3DMatch/3DLoMatch results of YOHO/RoReg/other baselines and all the origin version without adding random rotations. We also report registration performances of YOHO and other baselines on rotated version 3DMatch and 3DLoMatch in Table.1.

Yours,