Modified code for exporting eval pcd

[dongjae0107]

Hi :)

How about slightly modifying the logic in 'export_eval_pcd.cpp' that marks 'gt_point' as dynamic? Currently, with a voxel size of 0.1, if the 'nearest_point' in 'et_cloud' is farther than 0.05 (half of the voxel size), 'gt_point' is classified as dynamic. However, when using the voxel downsampling method, which marks the space that the voxel spans, a more accurate approach might be to check if 'gt_point' falls within the space that 'nearest_point' spans before marking gt_point as dynamic. This would improve the accuracy of the classification.

[Kin-Zhang]

I see. Thanks for your pull request.

Your way is more on voxel-way using the grid bounding to check where the original one is mainly based on the nearest neighbor distance. I may prefer the latter one still since some methods can save at point-level, and voxel-way will inflate the score in some way.

Maybe we need comparison results to show the benchmark methods' score change. Could you also attache a table for that if it's possible?

[dongjae0107]

I evaluated the performance of the benchmark methods on the KITTI 00 and 05 sequences using the approach I suggested. (I used GT poses, not SuMa poses.) While the performance rankings between the methods remain consistent, I observed slight changes in the SA and DA values.

I agree with you that the voxel-based approach may inflate the scores. I also think that using the original scoring method is more accurate for point-level methods. The voxel-based approach I mentioned would be appropriate for methods that classify dynamic elements at the voxel level. It would be a good option to be able to choose the scoring method depending on the method.