search

yccyenchicheng / AutoSDF

233 stars 29 forks source link

Question about the inferred mesh scaling factor #6

Closed AlbertoRemus closed 2 years ago

AlbertoRemus commented 2 years ago

Hello, I find this work really interesting.

I would like to ask a question about the scaling factor of the reconstructed 3D mesh: in this framework is this model reconstructed up-to-scale or it is capable to retrieve it with the correct scaling factor?

yccyenchicheng commented 2 years ago

Hi,

I am not sure what you mean by up-to-scale. In the preprocessing step for computing the T-SDF, we normalize the meshes (zero center and in unit-cube) following the DISN repo in this function: https://github.com/Xharlie/DISN/blob/master/preprocessing/create_point_sdf_grid.py#L169-L198.

AlbertoRemus commented 2 years ago

Hi @yccyenchicheng, thanks for your reply, I was asking: given an RGB image is the method capable of reconstructing the 3D model with the correct scale or not? To be more clear is the scale of the estimated 3D model close enough to the ground truth one?

yccyenchicheng commented 2 years ago

Hi @AlbertoRemus, thank you for the explanation!

I think it can't since we normalize all the meshes during training. So it predicts the 3D model at the normalized scale.

AlbertoRemus commented 2 years ago

@yccyenchicheng thanks for your answer! Hence, I'm gonna close the issue having the information I needed