I was wondering if the normal estimation code is available (even if some files cannot be run)? I am having difficulties in reproducing the normal estimation results in table 4 using PointNet and PointNet++. The best mean cosine loss for pointnet I can get is about 0.65, which is far from 0.47 in the table so I think I must have some dumb mistakes. Table 4 in your PCNN paper is below:
Here are some follow-up questions:
I was wondering if you have any advice or any code for normal estimation part I can read?
Also, about the cosine loss in the table, is the cosine loss defined as 1 - cos_similarity?
About the pointnet part segmentation code, did you use the part segmentation structure in the pointnet main paper (i.e. 1088-channel before FC layers) or did you use the enhanced version of part segmentation structure mentioned in the pointnet supplementary material (i.e. 3024-channel before FC layers)? I have pasted the two structures below:
The original Pointnet part segmentation structure:
The enhanced version part segmentation structure, mentioned in pointnet supplementary:
Hi @matanatz,
Many thanks for releasing the code.
I was wondering if the normal estimation code is available (even if some files cannot be run)? I am having difficulties in reproducing the normal estimation results in table 4 using PointNet and PointNet++. The best mean cosine loss for pointnet I can get is about 0.65, which is far from 0.47 in the table so I think I must have some dumb mistakes. Table 4 in your PCNN paper is below:
Here are some follow-up questions:
cosine lossdefined as1 - cos_similarity?The original Pointnet part segmentation structure:
The enhanced version part segmentation structure, mentioned in pointnet supplementary:
Best, Zirui