Florian Echtler

Oh, now I know what you mean... I guess that would involve improving the training set and re-training the CNN. Did you try any of the other models, in some...

This is pretty easy to test, I just pushed 5660433 which uses the standard OpenCV bilateral filter before handing the image to the CNN. I don't see a lot of...

@BenBE OK to close this issue?

[offtopic] +1 for "hipster tooling" ;-) [/offtopic]

Thanks for the heads-up. I wanted to try Twitch anyway sometime, I'll let you know when I get around to test this.

Hmmm, interesting idea. That might work with deeplab, worth a try. For the Google Meet model, I think it's mostly trained on images showing the full portrait in the center...

Good point, this should be easy to try.

This is already happening here: https://github.com/floe/deepbacksub/blob/main/deepseg.cc#L303-L306 Probably needs a bit of tweaking on size of `info.element`...

Hm, I had considered finding the outer contour and filling everything inside, but that would fail with e.g. raised arm touching the head?

That's going to be rather difficult, because you need to split it into small chunks of at most 21 bytes, and there's no guarantee that any individual chunk will actually...