search

1zb / 3DILG

3DILG: Irregular Latent Grids for 3D Generative Modeling
https://1zb.github.io/3DILG/
Other
88 stars 6 forks source link

normalization? #5

Closed xwhan closed 2 years ago

xwhan commented 2 years ago

Hi @1zb,

I have another question regarding to data processing? Does the model expect normalized point cloud, SDF coordinates (e.g., transform all points to be within [-0.5, 0.5] ^ 3)? It seems that you scaled the points cloud but not the SDF points? https://github.com/1zb/3DILG/blob/master/shapenet.py#L134

I'm also curious why the "Axis transform" is only applied on training data. Is this a common practice? -- sorry that I'm new to this line of work :) https://github.com/1zb/3DILG/blob/master/datasets.py#L33

Thanks in advance!

1zb commented 2 years ago

Does the model expect normalized point cloud, SDF coordinates (e.g., transform all points to be within [-0.5, 0.5] ^ 3)?

You are correct about this. Points are scaled to [-1, 1] in our experiments. However, other scales should work too, e.g., [-0.5, 0.5].

It seems that you scaled the points cloud but not the SDF points?

We generate SDF by ourselves. They are already scaled.

I'm also curious why the "Axis transform" is only applied on training data. Is this a common practice? -- sorry that I'm new to this line of work :)

Axis transform is a data augmentation technique. We do not need it when testing.

Best, Biao

supersyz commented 1 month ago

Hi, Congratulations on your great work! Would you mind tell me how you calculate the scale values from xxx.npy,which is applied to the point clouds? https://github.com/1zb/3DILG/blob/master/shapenet.py#L134