virtual-background
virtual-background copied to clipboard
Segment Anything Version
Any thoughts on using Segment Anything? Seems like it might be a good fit to replace the Google models.
+1
Sounds interesting indeed. For now I don't fully understand the best way to provide Segment Anything with the class "person" or "body" that we'd like to segment. Needs a bit of investigation and experiment.
Grounded segment anything with "Human" as input seems to work pretty well.

But I do not really know much in Python universe so I don't know the struggle it would be to add this into the project
https://twitter.com/xenovacom/status/1650634015060156420
With Onnx getting WebGPU support and Segment Anything offering an Onyx model, should be possible starting with Chrome 113.
Also relevant: https://github.com/gaomingqi/Track-Anything