Questions about training/testings set

[ziruiw-dev]

Hi @sitzikbs,

Thanks for this work and releasing code. I have two questions regards your paper.

1. In section 4.1 Datasets, you mentioned the PCPNet test dataset has 22 shapes. I downloaded the PCPNet dataset from the PCPNet project page and found there are only 19 shapes in the test set, and 3 shapes in the validation set and the validation set is a part of the test set, so if I understand correctly, there are only 19 shapes in the test set?

2. In section 4.2 Training details, you mentioned you trained with 32768 (1024 samples x 32 shapes) random subsets. Does that mean you count 8 shapes with 4 noise levels as 32 shapes?

Best, Zirui

[sitzikbs]

We used exactly the same dataset and train/test split and setup as PCPNet.

To answer your questions:

1. You are correct. only 19 shapes. 15 + 3 analytic +1 boxy_smooth, just like in PCPNet.
2. Yes, 8 shapes with 4 noise levels = 32 shapes.

[ziruiw-dev]

Hi @sitzikbs, Thanks for your reply. Best, Zirui

[ziruiw-dev]

A follow-up question, I noticed that you train with noise level 0.00125, 0.006 and 0.012 and I think none of them exists in the PCPNet dataset. May I ask did you augment those noise levels by yourself?

[sitzikbs]

Which file are you referring to?

I trained with the same data they provided. I think these values correspond to their low, medium and high noise levels. All are a percentage of the bounding box.

[ziruiw-dev]

Thanks for replying.

When I look at the PCPNet dataset, for one shape, the xyz data in their dataset look like this:

• noise-free
  • armadillo100k.xyz
• gradient noise
  • armadillo100k_ddist_minmax.xyz
• striped noise
  • armadillo100k_ddist_minmax_layers.xyz
• brown noise
  • armadillo100k_noise_brown_1.00e-01.xyz
  • armadillo100k_noise_brown_1.00e-02.xyz
  • armadillo100k_noise_brown_2.00e-02.xyz
  • ...
• white noise
  • armadillo100k_noise_white_1.00e-02.xyz
  • armadillo100k_noise_white_2.00e-02.xyz
  • armadillo100k_noise_white_3.00e-02.xyz
  • ...

so I am not sure what file do you refer to when your use noise standard deviation sigma = 0.00125 or sigma = 0.006? The first image below is from your Nesti-Net paper, the second image is the PCPNet dataset.

Many thanks.

[ziruiw-dev]

Ok... I think I understood what you mean. The PCPNet dataset provides a trainset_whitenoise.txt file that specifies they trained with white noise level 1e-2, 5e-2, 1e-1.

Thanks a lot.