search

orangeduck / lafan1-resolved

Resolving of the Ubisoft La Forge Animation dataset onto a common skeleton
Other
30 stars 1 forks source link

How to convert your bvh to smpl? #1

Open GuoweiXu368 opened 1 month ago

GuoweiXu368 commented 1 month ago

Great work! Do you know how to convert your bvh to smpl?

Shimingyi commented 4 weeks ago

In my experiences, there are several ways to do it.

  1. You can try joints2smpl, which fits SMPL parameters using joint positional constraints and the vposer prior to ensure realism for each frame. However, this method is quite slow, and the quality of the fitting - especially in the temporal coherence, may not meet expectations.
  2. Another method first obtains an aligned skeleton in SMPL by setting all shape parameters to zero and fitting only the shape parameters. This will result in a SMPL-T-Pose skeleton, and you can convert it into bvh file. Then you can retarget other bvh assets by using software like MotionBuilder or Blender to create new bvhs with the desired SMPL-T-Pose. The final step is to convert the Euler angles in the new bvhs to the axis-angle representation in SMPL and store them in the numpy file as same as SMPL. This way can generally give better performance and flexibility on the motion quality.
GuoweiXu368 commented 2 weeks ago

In my experiences, there are several ways to do it.

  1. You can try joints2smpl, which fits SMPL parameters using joint positional constraints and the vposer prior to ensure realism for each frame. However, this method is quite slow, and the quality of the fitting - especially in the temporal coherence, may not meet expectations.
  2. Another method first obtains an aligned skeleton in SMPL by setting all shape parameters to zero and fitting only the shape parameters. This will result in a SMPL-T-Pose skeleton, and you can convert it into bvh file. Then you can retarget other bvh assets by using software like MotionBuilder or Blender to create new bvhs with the desired SMPL-T-Pose. The final step is to convert the Euler angles in the new bvhs to the axis-angle representation in SMPL and store them in the numpy file as same as SMPL. This way can generally give better performance and flexibility on the motion quality.

Wow, thank you so much for sharing your valuable experience—it's incredibly useful!