The collision method of Simplicits

[SCUTykLin]

Hello:
Thank you very much for open-sourcing such an influential work.
Could you please tell me what collision detection method you use? Is it IPC[1]? I noticed that you cited this work in your paper.

[1] Incremental Potential Contact: Intersection- and Inversion-free, Large-Deformation Dynamics

[itsvismay]

Hello,

For detection, we use warp's hash grid structure. You can find out more about it here.

Once collisions are detected, we use an IPC type barrier energy to resolve collisions between points in two different objects

[SCUTykLin]

Hello,

For detection, we use warp's hash grid structure. You can find out more about it here.

Once collisions are detected, we use an IPC type barrier energy to resolve collisions between points in two different objects

Thanks for your reply. I see that what you mean.
And anthor question: I'm quite puzzled as to why adding an Orthogonality loss can prevent the model from getting stuck in local optima. Could you please provide the idea or origin behind this loss?

[itsvismay]

We want the weights to be linearly independent, and orthogonality is a stricter condition than necessary, but a simple loss function.

Without linear independence, the simplest way to minimize elastic energy is to collapse all weights 0.

[SCUTykLin]

We want the weights to be linearly independent, and orthogonality is a stricter condition than necessary, but a simple loss function.

Without linear independence, the simplest way to minimize elastic energy is to collapse all weights 0.

Thanks for your reply. If u want the weights to be independent, why not fourier embedding? e.g,. embedding used in Nerf

[itsvismay]

Fourier features/positional encoding are a good idea, but they are useful for mapping higher frequencies.

In our case, we want to capture the low-frequency deformation bases of the object