ethz-asl / 3dsnet Goto Github PK

Hi, Thanks for nice works and code.

I have a question about installing torchvision.

Why do you install torchvision of "conda install pytorch torchvision cudatoolkit=10.1 -c pytorch --yes" and "pip install git+https://github.com/rtqichen/torchdiffeq torchvision"?

What is the difference between the two?

Hi, Thanks for your wonderful code and it's helpful for me.
But, I feel uncertain in detail when conduct experiments:
In your paper 3DSnet subsection 4.4, it said that the number of armchair and straight chair is 1995 and 1974 respectively while i got the size are 794 and 1240 in practice. （I download the experiment data from the link offered in download_shapenet_pointclouds.sh)
Wish your reply soon!

Why do I show 'No module named 'chamfer_3D''

Hi,

I'm wondering since Pytorch3d only accepts Python>=3.8 (https://github.com/facebookresearch/pytorch3d/blob/main/INSTALL.md) and Pymesh only supports Python<3.7, how do you find a matching version of all these packages?

Thank you!

Line 351/354 and line 359/362 should be norm2 and norm3?

Dear Mattia,

Sorry to bother you, but I'm having difficulties in understanding the demo.sh code. I guess demo.sh is to test our trained networks with our test dataset? Please correct me if I'm wrong.

The main part I don't understand is line 509 in trainer.py:
data_a = EasyDict(self.datasets.dataset_test[self.classes[0]][index_a]) (same for data_b)

I think data_a['points'] should be the normalised and downsampled shape, I can see it has 2500 points. However, I ran train.sh and demo.sh with --number_points = 642 and --decoder_type = 'atlasnet'. I really don't know why in demo.sh, each loaded sample has 2500 points but not 642 points.

Thank you in advance. Looking forward to your help!

Sincerely,
Wei