search

Yzmblog / SurfD

Surf-D: Generating High-Quality Surfaces of Arbitrary Topologies Using Diffusion Models (ECCV 2024)
154 stars 6 forks source link

data processing #5

Open gzgwink opened 3 months ago

gzgwink commented 3 months ago

Thank you for open sourcing such excellent work, I have a question about data set processing, is it necessary to normalize obj to the range [0,1]? Why not normalize to [-1,1]. https://github.com/Yzmblog/SurfD/issues/3#issuecomment-2227612152 And the code mentioned in the figure below is to move the center point of the [-1,1] range obj to (0,0,0). Is this also necessary when processing with dataset? image So my main doubt is the scope of normalization.

gzgwink commented 3 months ago

image During data processing, is clipped values[0,0.1] linearly normalized to [0,1]? Why is this step removed from the code?

Yzmblog commented 3 months ago

Hi, Thanks for your interests. For your question: 1. Move the obj center to (0, 0, 0) and normalized to [-1, 1] is correct. 2. For the reusability, these codes are moved to the training script, the udf value are still clipped and normalized to [0, 1] when training.