Basic info
My code of pipeline of paper: SuperTrack: Motion Tracking for Physically Simulated Characters using Supervised Learning
I use LAFAN1 as my database which you can get from LAFAN1.
Paramaters presentataions in this repo
- world space positions
$x^p\in \mathbb{R}^{3B}$ - velocities
$\dot{x}^p\in \mathbb{R}^{3B}$ - accelerations
$\ddot{x}^p\in \mathbb{R}^{3B}$ - rotations
$x^r\in \mathbb{R}^{4B}$ : in quaternions - rotational velocities
$\dot{x}^r\in \mathbb{R}^{3B}$ : in an axis-angle vector and convert into rotational matrix by exp map. - rotational accelerations
$\ddot{x}^r\in \mathbb{R}^{3B}$ : in an axis-angle vector to be integrated and just add to$\dot{x}^r$
Reference
There is other information that may help you reproduce this paper from scratch.
You can learn rotational velocities and rotational accelerations from Section 4.1 of Orientation, Rotation, Velocity and Acceleration, and the SRM and Rigid Body Kinematics.
Motion Capture File Formats Explained explains what is bvh file clearly.
On the Continuity of Rotation Representations in Neural Networks explains why convert quaternions into the two-axis rotation matrix format.
Demo
-
ground truth
-
pred
Due to my uncertainty regarding how to check my code, there may be some errors here. Please use with caution.