Regarding the improvement of structural completion effect.

[Jueqian]

Sorry to bother you again, but I have a few questions would like to ask.
I noticed that the originally flat ground becomes bumpy in the completion results, as well as the walls, and the details of foreground objects like vehicles are not very clear. I want to achieve a more refined structural effect. Do you have any suggestions? Would it be easier for me to focus on the refinement network in the second stage?

[nuneslu]

Hi, my first suggestion would be to check the generated ground truth. One known problem from the KITTI dataset is the ground truth poses. The provided poses are not 100% accurate and can lead to some alignment, such as objects not correctly aligned during the point cloud aggregation. Also, the point cloud is not deskewed (check here for more info), which can increase this alignment problem even more. If you manage to have a more accurate ground truth, the results should also get better. Besides that, you could try improving the refinement network or use some more recent diffusion regularization, such as SNR loss weighting. All the individual parts of the pipeline can be improved, but I guess that checking the ground truth data and using more recent diffusion regularization/training schemes could lead to more improvements.

[Jueqian]

Your answer is fantastic! Thank you very much for your insights. Wishing you all the best!