inference results export to bvh error

[Italian-PAO]

Hello, thank you very much for interGen's contribution to this field!

I am using the action results (3d pose) generated by interGen reasoning to redirect some fbx models. My approach is to use the conversion tool provided in momask (https://github.com/EricGuo5513/momask-codes/blob/main/visualization/joints2bvh.py) to convert the 3d pose into a bvh format action, and then use bvh for redirection.

However, I found some problems in the process of converting bvh:
When one person in a two-person action is facing forward, the converted BVH action is very correct; but when one person is facing away/sideways, the converted BVH action is very strange, but the visualization of 3D pose is normal (see the two videos in the attachment).

I suspect that there is a problem with the IK solution during the conversion process, but I am not an expert in this area, so I would like to ask you:
What might be the problem in this case? If it is really a problem with the IK solution, how to solve this problem?

Maybe this issue is not appropriate to raise under your repo, but I still want to ask for your help, thank you very much!

Looking forward to your reply

大佬你好，非常感谢interGen为这个领域做出的贡献！

我正在用interGen推理生成的动作结果（3d pose）来重定向一些fbx模型，我的做法是使用momask中提供的转换工具 (https://github.com/EricGuo5513/momask-codes/blob/main/visualization/joints2bvh.py) 将3d pose转换成bvh格式的动作，再用bvh来做重定向。

但是在转换bvh的过程中我发现了一些问题：
当双人动作中某一个人的朝向是正对着的时候，转换出的bvh动作是非常正确的；但当某个人是背对/侧对着的时候，转换出的bvh动作就非常奇怪，但3d pose的可视化却是正常的（可以查看附件的两个视频）。

我怀疑是转换的过程中IK解算出现了问题，但我并不是这方面的专家，所以想要请教您一下：
这种情况可能是什么问题？如果真的是IK解算的问题，该如何解决这个问题？

可能这个issue在您的repo下面提并不合适，但还是想请求您的帮助，非常感谢！

期待您的回复

Two.people.bow.to.each.other.mp4

exported.bvh.mp4

[tr3e]

ik优化不鲁棒导致的吧，试下分阶段优化，先优化躯干位置，再优化四肢

[Italian-PAO]

非常感谢大佬的回复！我是这方面的小白，想请教一下大佬有没有可以参考的分阶段优化代码呢？再次感谢！