Linear blend skinning

[MaxMayya]

How do you initialize W weights? What loss do you use to train the W layer?

[jby1993]

We constrain the weights of each vertex to be non-zero only on its nearby several parts and require the sum of these parts' weights equals one. Actually, I do not remember whether I does the initialization. You can initialize these weights based on the geodesic distance of the vertex to related parts. I think the initialization is not important if the abovementioned two constraints are satisfied.

[MaxMayya]

Thanks for the reply. Yes I thought of using the geodesic distance but I cannot figure out how to impose those constrains. Correct me if I'm wrong: in the code, the weights of W (loaded from RepWs.pth) seem not to satisfy the convexity constrain (sum of positive weights to 1). Some weights are negative and the forward of learnable_skinning_layer normalises them through a SoftMax. I'm a little puzzled here!

[jby1993]

Yes, I use a softmax to normalize the related parts' weights of RepWs to satisfy the convexity constrain. The softmaxed weights are the final skinning weights.

[MaxMayya]

Okay I see. What about the sparsity constraint? Do you fix the number of the "influencing" parts a priori or it's automatically done during the training? If it's the latter, how do you guide the learning to satisfy such constraint?

[jby1993]

I fix the number of the influencing parts for each vertex, and require the convexity constrain on these influencing parts.

[MaxMayya]

So when fixing that number, do you follow some semantic logic (for ex. hand vertices can be influenced by forearm and arm but not by the feet)?

[jby1993]

yeah

[MaxMayya]

thanks