leoshine / Spherical_Regression

PyTorch implementation of cvpr2019 paper "Spherical Regression: Learning Viewpoints, Surface Normals and 3D Rotations on n-Spheres".
Other
147 stars 17 forks source link

Questions about the construction of ModelNet10-SO(3) dataset #10

Closed yd-yin closed 3 years ago

yd-yin commented 3 years ago

Hi, thanks a lot for your great work and useful dataset!

Could you elaborate a bit more about the construction of the ModelNet10-SO(3) dataset? since I may want to reconstruct and re-render the images

For example, what are the intrinsic and extrinsic parameters of the camera? Is the original shape in ModelNet10 the reference "canonical" pose (i.e., identity rotation)?

Looking forward to your response Thanks!

leoshine commented 3 years ago

Hi,

Thanks for the interests of the dataset.

For your convenience, one example of intrinsic matrix (OpenGL) based on above setup can be:

[[  5.8   0.    0.    0. ]
 [  0.    5.8   0.    0. ]
 [  0.    0.   -6.  -17.5]
 [  0.    0.   -1.    0. ]]
yd-yin commented 3 years ago

Hi, Thanks for your quick response and informative message!

I finally found that maybe it will be more convenient to render by myself:)

Thanks!

yd-yin commented 3 years ago

Well, sorry for the bother, but I got another problem...

I tried to render 3D models from ModelNet10 dataset by trimesh + pyrender, but just to find the renderings of some models are incorrect.

I'm not sure if it is caused by wrong face normals. Did you meet the same issue before?

Thanks a lot!

b360a44117afe479329cefe759e7b5e

leoshine commented 3 years ago

Indeed. It can be the wrong surface normal or wrong winding of the three vertices of a face (if Trimesh generates the surface normal automatically). I ended up generating the surface from the triangle face definition myself, instead of using any surface normal defined in the CAD file. This at least corrects the back-facing problem, however, the rendered surface may not be that smooth.