Some questions about test dataset about idam HOT 6 CLOSED

Hello,

Thanks for your interest!

1. It's been a long time ago and I do not remember every detail of the code. The data.py file is borrowed from DCP and PRNet, but is modified many times so it may not be exactly like the original. If you want to check if the transformation is fixed, you can print out immediate results of the first entry of the dataset to see if it changes every time.

2. PRNet uses DGCNN, which computes distance in feature space, so it is dynamic. GNN is a very general term, and covers a lot of architectures. I used a very naive version of it. The DGCNN paper lists some GNN variants, if you are interested.

Best,
Jiahao

from idam.

Thank you for your reply. Do you have any pre-trained checkpoints for unseen shapes in ModelNet40 with Gaussian noise?

from idam.

Sorry we do not. But training a new one should be very quick.

from idam.

OK, thank you again for your prompt reply!

from idam.

Dear authors,
Why does the use of other features to embed the network make a big difference between the training results and the test results? such as a pointnet or DGCNN.
The result of using the author's GNN will be normal.

from idam.

Dear authors,
Why does the use of other features to embed the network make a big difference between the training results and the test results? such as a pointnet or DGCNN.
The result of using the author's GNN will be normal.

I am not sure about that, I have not tried any larger network. Is it possible that the network is over fitting the dataset? Maybe you can reduce the number of parameters and try again.

from idam.