Zed 2 human volumetric capture prototype

@olivierbraun led dev

Hello world
We are trying to do a volumetric capture of people using a zed 2 cam in unity
latest Sdk 3.5.0 and Cuda 11
The segmentation mask is not very accurate and we have a lot of artefacts and proximity distortions .( we have the same problem in all resolutions modes )
I understand that this is not a basic use case for this camera :slight_smile:
My question
Is it possible to enhance the segmentation mask in order to match at least the quality of the kinect v2 user index map ?
Any help is appreciated

Even if there is depth distortion, I would think one can use the per pixel depth to isolate a cluster of pixel sharing similar depth values (human body) versus the rest of the scene pixel, no?
If you have an AI segmentation mask, that should give you an even better estimate of the median depth of the human body you want to isolate.
The only tricky part I think would be to differentiate the feet from the ground accurately or any piece of the background which is touching the human body, like a chair or a table for instance.