After more testing, I found that hardware choice mattered more than I expected. Using a depth camera optimized for surface stability helped more than aggressive post-filtering.
In my setup, switching to revopoint surface pro 50 reduced Z-axis jitter on semi-reflective metal and coated plastic parts, even without matting spray. Frame-to-frame depth consistency was noticeably better, which made temporal averaging more effective.
It doesn’t fully eliminate reflection artifacts, but compared to a generic structured-light module, the depth data was cleaner in real mobile robot motion.
Sharing this in case it helps others dealing with similar issues.

 

What is very interesting is the actual concept of such noise. Of course, I am more conversant with teleconference cameras and they are not nearly that perceptive.
But the most amazing cameras of all are the ones used for sensing on computer driven cars.. Those cameras need to sense distance accurately at all times, and the software associated with them needs to recognize humans and their motion almost instantly under all conditions. And do all of that repeatedly without daily checking and cleaning .
What sort of application are you using the cameras for ?? Presently my applications are hobby entertainment. Not serious yet.