search

DeepMotionEditing / deep-motion-editing

An end-to-end library for editing and rendering motion of 3D characters with deep learning [SIGGRAPH 2020]
BSD 2-Clause "Simplified" License
1.57k stars 257 forks source link

Retargeting retraining on customized dataset #210

Open HaozhouPang opened 2 years ago

HaozhouPang commented 2 years ago

Hello Peizhuo,

Thanks for making this amazing project public available!

I am trying to retrain the retargeting model on my own dataset, I have a new skeleton and its motion data. As your suggestion said the motion data should have the same T pose, does it mean the bvh files should start from the same T pose?

Thanks, Haozhou

PeizhuoLi commented 2 years ago

Thanks for your question. It doesn’t have to be exactly the same T-pose, but should be similar. Because in our experiment, drastically different T-poses may introduce extra burden to the network and eventually can fail it (in fig 15). In the pre-trained model, the T-poses in the dataset are similar but not the same.

Hope this can be helpful.

HaozhouPang commented 2 years ago

Thanks for your reply!

If I understand this correctly, the T-pose of different characters should be similar to get better result, but the bvh animations that used for trainning and inference does not need to start from T pose in the first frame, right?

Best, Haozhou

PeizhuoLi commented 2 years ago

You are correct!