Peizhuo Li

Hi, they are the same thing but in different representations. Precisely speaking, our pose file is the joint rotation in rotation vector representation with the order of the name of...

I use this representation here for a very simple reason: the dataset I used for training accepts this representation for joint rotations : ). In other part of the code...

Hi, as far as I know, this is still an open problem. First is we are only able to calculate skinning weights for a T-posed mesh. Second is that neural...

Thanks for your great question! In general cases, each blend shape corresponds to some semantic meaning (smile, sad, etc.). But here we extend it such that the residual deformation is...

Hi, theoretically speaking it is possible to rig fingers though we haven't tried yet. The method itself naturally supports fingers providing a feasible training dataset. But fingers can introduce a...

Hi, the training code has been released. Thanks for your interest!

Hi, the dataset we use doesn't contain finger's rigging. One possible solution could be switch to [SMPL-X](https://smpl-x.is.tue.mpg.de) model.

您好，类似的情况我们在实验中是可以解决的 (如论文中的figure 12)。intra的retargeting应该是相对容易的，通常只要reconstruction loss收敛就可以达到比较理想的效果。如果intra效果不好的话不妨检查一下reconstruction loss方面的问题。

据我所知数据不足的问题目前还没有比较好的解决方法；不过如果训练数据的动作能较好的覆盖测试集的话效果应该不会太糟糕

我认为你举的这个例子是不一定可以成功的，需要双向的数据都有一些才会比较容易成功