Closed Junlin-Yin closed 3 months ago
You need to divide accscale for AMASS dataset. I use the same scale for different datasets. Maybe your (DIP) method can also work.
I have a few questions that I'm confused about: 1.What are the functions of acc scale and vel scale?
I just experimentally found that the max/min value of accelerations in all the datasets is around 30m/s^2 (or 90m/s^2?), so I scale it by 1/30 to ensure similar scale w.r.t the orientation, which is in [-1, 1]
The velocity scale is similar to acceleration scale. The 1/60 is delta t between frames. The output velocities should also be rescaled in around [-3, 3], so I use a vel_scale.
Thank you very much for your answer
Hi, I'm currently trying to reproduce your project, and I find
normalize_and_concat
function a little bit confusing. You just divide the normalized acceleration byacc_scale
which is 30, and this function is applied to both DIP-IMU and TotalCapture. The problem is, do I need to divideacc_scale
for AMASS dataset as well during pre-training?Moreover, I find that DIP project simply uses the z-score standardization to "align" different datasets, which is given up in your project. Therefore, I speculate that dividing acceleration by
acc_scale
is your way to standardize or align AMASS, DIP-IMU and TotalCapture. Am I right?