Hello Mohamed,
i hope you are doing well. Thanks for your interesting and insightful benchmarking effort! It would be great if you could answer a few questions regarding details of your evaluation setup:
-
On p. 7 you write: 'In this work, all models are trained on sampled point clouds with 10,000 surface points.' On first glance, this number seems comparatively low, but I understand that some of your baseline methods are limited in the amount of points they can process at once.
- Did you find that performance did not increase for larger point clouds?
- Did you perform random subsampling or something like farthest point sampling?
- Also, i was a bit confused about this: 'To assess robustness beyond training resolution, we
evaluate predictions on both this 10,000 subsampled mesh and the full-resolution surface mesh containing 487,846 nodes.' I think the number of nodes varies between different test samples?
-
I assume that metrics are computed per sample and then averaged across samples? Alternatively metrics could be computed based on all concatenated points at once.
Best,
Jan
Hello Mohamed,
i hope you are doing well. Thanks for your interesting and insightful benchmarking effort! It would be great if you could answer a few questions regarding details of your evaluation setup:
On p. 7 you write: 'In this work, all models are trained on sampled point clouds with 10,000 surface points.' On first glance, this number seems comparatively low, but I understand that some of your baseline methods are limited in the amount of points they can process at once.
evaluate predictions on both this 10,000 subsampled mesh and the full-resolution surface mesh containing 487,846 nodes.' I think the number of nodes varies between different test samples?
I assume that metrics are computed per sample and then averaged across samples? Alternatively metrics could be computed based on all concatenated points at once.
Best,
Jan