Hi, I'm having some problems with the reproduction training process and hope I can get some help from you.

1、Regarding the rotation information we used in the DIP-IMU training and the position information we synthesized subsequently. Could we just use the “gt” data in the DIP-IMU dataset to synthesize all the data we use for training, regardless of “imu_ori” and “imu_acc”?
2、The DIP-IMU dataset doesn't have translation data, so can we train Trans-B1(Va) and Trans-B2(Cf)? Or is it only used to train the first three networks?
3、You said "Maybe you can check the estimated velocity v.s. the ground truth" in (#11), isn't it true that we can just treat the position difference as velocity without divide by the time interval delta_t? From my training results, I find that divide by delta_t doesn't rather reasonable for the training result.
4、You said in (Xinyu-Yi/TransPose#65) that the joint positions generated by forward kinematics in DIP-IMU and TotalCapture need to be rotated by left-multiplying the inverse root rotation, so do we need to do the same for the joint.pt generated from the AMASS dataset in order to train it?