Hello,
I would like to report a behavior I’ve observed when using the 18-point and 34-point body models with the ZED SDK.
-
With the 18-point model, I’m unable to obtain reliable segment orientations (it seems rotation information is missing or insufficient).
-
With the 34-point model, segment orientations are available, which is very useful. However, I observe an obvious low-frequency noise/drift (a slowly oscillating offset over time).
For example, the ear markers tend to drift upward and can end up on the top of the head, which does not happen with the 18-point model.
My assumption is that this behavior may be related to the body fitting process and possibly the inverse kinematics used in the 34-point model.
I have a few questions:
-
Is this a known issue?
-
Is there a clear explanation for this behavior (differences in pipeline between the 18- and 34-point models, fitting, filtering, etc.)?
-
Are there any recommended settings or workarounds to reduce this effect for the 34 points model (SDK parameters, smoothing options, etc.)?
I can also provide videos demonstrating the issue, if that helps with the diagnosis (I can send them by email if needed).
Thank you in advance for your help and insights.