[Feature Enhancement] Motion retargeting from OpenPose JSON output about deep-motion-editing HOT 4 OPEN

Thank you for your interest. Unfortunately, converting joint position (json_output_from_openpose) to joint rotation (bvh) is not a trivial task and the latter representation is necessary for driving skinned model.

Although reconstructing 3D animation from monocular video is beyond the scope of this library, it is a popular research area and there are some nice works. You could try to integrate other projects and train your own model.

from deep-motion-editing.

Thanks for the definitive answer, I have 1 logical question though, might be a very beginner-mindset question.

Can't we use relative rotations / movements in an animation (one .bvh which contains one type of armature doing something) and impose those relative notations onto a different armature (one .fbx which contains a skinned model with its own armature, but no animation) without using a trained model?

from deep-motion-editing.

I'm not quite sure what does relative rotation mean. But I think that's exactly what motion retargeting does: transfer an animation from one armature to another different armature. And of course, there exists related works without using deep learning method and possibly can deal with unseen armature.

from deep-motion-editing.

By relative rotation I meant, for example relative rotations and movements in parent-child structures, so effectively we'll have a core bone, and other bones' relative rotations to represent the whole structure, and how this would help is, it would standardize over almost any armature type. (Just a vague beginner idea, could be completely wrong here).

Apart from that, would you be kind enough to point me to some other existing related works? I tried exploring, but in vain.
Thanks anyways. 👍

from deep-motion-editing.