DGR vs FCGF

[Solonets]

Hi,

Thank you for the interesting work.

In table 3 of the paper, you compare DGR and FCGF in terms of speed and accuracy. I wonder how DGR can be quicker then FCGF. As far as I understood DRG pipeline includes evaluation of FCGF as one of the steps, so the time should be higher for the DGR. Maybe these times stand for something else. Can you please clarify that? Also, it looks like DGR performed worse than FCGF in the first place. Isn't it better just to use FCGF followed by Ransac instead?

Thank you for your answer.

[chrischoy]

I see that the table is not very self-explanatory.

FCGF by itself is not a registration algorithm.

We used the RANSAC like the other experiments for FCGF and it is very slow when the number of voxel is large.

Specifically, for KITTI like outdoor lidar scans, there are so many voxels that RANSAC becomes very expensive.

Instead, DGR uses one feed forward pass with 1 nearest neighbor which is a lot faster.

[Solonets]

Thank you for your answer. It's clear now.

I also wounded if you compared FCGF + RANSAC+ICP vs DGR + ICP. I feel that would be an expository comparison.

[XuyangBai]

Hi @chrischoy, thanks for your sharing.

I have a related question for Table 3. Why is the computation time for FCGF is much longer than RANSAC (6.38s vs. 1.37s)? I assume for RANSAC you are using fpfh as the descriptor. So is it because extracting FCGF is slower than extracting fpfh from the input point clouds? Or it's because the correspondence set built by FCGF is much larger than that built by fpfh after the reciprocity tests?

[chrischoy]

Point clouds used in KITTI are much larger than pointclouds used in 3DMatch. There's no reciprocity test in FCGF.

[XuyangBai]

Hi @chrischoy I mean the RANSAC and FCGF in Table 3, aren't they both experiments on KITTI?