Point Cloud Pro tip: use a rolling average inside a double buffer to average pixel positions to avoid bouncing animations

give it a try:

A general... Higher dimensional noise tip frankly...

What model is this? Results look great

Thank you! Actually it's a pretty old (not temporal stable) but metric model, ZoeDepth. You can see the same model without the rolling average here

Thinking this would actually be a really good self supervised metrics for training video depth models.

Was that used in the Rocky example? I noticed that bounced a lot

No, that was just a 1:1 with the depth data. Give it a try here

oh nice, i tried that to fix jitter and it totally smoothed things out, though i had to tweak the buffer size to get it right 🤔

oh i hit bouncing points too, switching to a simple rolling average in the compute pass fixed it for me. curious if you're using a separate buffer for the averages or doing it inline?

why not just try a video depth estimation model e.g. video depth anything to ensure consistency? If there is camera movement, just try Pi3/DA3

this experiment is from back 2022

My timeline doesn't go that far, but funded a link to this other post

rolling average in compute vs frame history in vertex — curious if the double buffer overhead matters more than the noise reduction for this scale. tried storing a velocity buffer instead for smoother damping?

Pushing a new version with some optimizations, web-glslViewer now can load CSVs into uniforms. In this case the min and max depth per everyframe