search

google-deepmind / Temporal-3D-Pose-Kinetics

Exploiting temporal context for 3D human pose estimation in the wild: 3D poses for the Kinetics dataset
http://arxiv.org/abs/1905.04266
Apache License 2.0
219 stars 27 forks source link

How to run it for our own dataset ? #1

Closed manansaxena closed 5 years ago

manansaxena commented 5 years ago

Hi, I loved the work you have done and its quite interesting. I had two doubts - 1)How to run it for our own dataset containing videos. 2)How to extract the final 3D-joint locations(after the bundle adjustments)?

cdoersch commented 5 years ago

Unfortunately, we are currently unable to release our original implementation due to its reliance on internal tools.

We include 3D keypoint locations in our released file (3d_keypoints). If you're re-implementing the algorithm, you will need to use the SMPL model to compute the 3D mesh vertices from the pose and shape vectors, and then regress to 3D keypoints using the supplied regression matrix. This joint regressor is implemented in the hmr codebase. See: https://github.com/akanazawa/hmr/blob/master/src/tf_smpl/batch_smpl.py

manansaxena commented 5 years ago

Thanks a lot for the help.