How can I improve point cloud quality?

The ones above are point clouds I acquired using the Depth Viewer (with the ZED 2i camera).

I’m going to use python for acquiring new point clouds but I was wondering how to reduce noise and get a better quality. The object I have to capture is shiny. It’s a metal object.
I was wondering how to tune parameters in order to get the best quality.
Should I try to use the neural depth mode?
Should I try to adjust the max distance?
chatGPT suggested me to filter the point cloud to reduce noise by using the model sklearn.cluster.
I’m currently using ULTRA depth mode, when I try to use the NEURAL mode I get an error (failed to optimize model). Should I try another PC with higher computational capability to see if is able to use NEURAL?
I tried to tune other parameters like confidence, texture confidence, depth stabilization, brightness, constrast, hue…
But I need some advice. Should I try to move the camera while using depth stabilization to better fuse different frames into one in order to get a picture which is less noisy?

Thank you in advance for your support.

Hi @Daniele,

Neural Depth mode should definitely give better results, at the price of needing heavier computation power. However, it should not fail to optimize the model, that looks like a bug. :thinking:

Can you share the result of ZED Diagnostic either here or to please? We’d like to investigate this with more context.

In the meanwhile, to have a better depth on static objects, you can indeed increase the value of depth_stabilization, it should give better results even with ULTRA. Changing the max distance should filter the environment depending on it, but not “improve” the depth per see.


The notebook I am using is old. It features a Nvidia GEFORCE GTX 950M (computational capability 5.0 in the list of the graphic cards). I’ll use a better PC for this work afterwords.