Thank you posting the code. I noticed that the results for Flownet3d shown in Table 1 of your paper do not match the ones shown in Table 2 & Table 4 of the original Flownet3d paper. Can you please comment on this?
When we submitted their paper, since flownet3d code & details is unavailable, we reimplemented a version based on their Arxiv submission at that time (mentioned in our paper).
Besides, some known differences are:
They sample 2048 points per point cloud while we sample 8192 points per point cloud.
They used "Inference with random re-sampling" strategy while we do not use it as one of our focus is on speed and this strategy slows down inference speed.
For KITTI, they used annotations for ground removal while we remove ground points by height since the ground removal annotations are not publicly available.
On large KITTI scenes, they split the scene into multiple chunks while we do not have this preprocessing step.
Hi
Thank you posting the code. I noticed that the results for Flownet3d shown in Table 1 of your paper do not match the ones shown in Table 2 & Table 4 of the original Flownet3d paper. Can you please comment on this?
Thanks Himangi