About vertex normal of object

[dqj5182]

First of all, I really enjoyed reading your work! It is a phenomenal work that hugely contributes to human-object interaction research.

My question is about the object normal n_o in Figure 2. I am not sure if you have encountered this issue, but some datasets (e.g., ShapeNet) has numerous inverted faces within 3D object mesh. In such cases, I found it difficult to use your method to learn accurate contact as object normal values are misleading.

I want to ask if you have also encountered into similar problem and if so how you dealt with the problem when extracting contact with object normal.

[jellyheadandrew]

Hi thanks for the question!

We use Poisson disk sampling implemented in the open3d framework (link) which outputs pseudo-equispaced point cloud extracted from the mesh surface.

For meshes with inverted faces, I'd need to check if open3d implementation handles that too. I'll be in touch soon

Thank you!

[dqj5182]

Thanks for your answer! I will check the open3d implementation.