Performance Idea - option to reprocess mask every X frames about obs-backgroundremoval

[iBenMoore]

Looks like there might be a Mediapipe meet onnx model here to try: https://reposhub.com/dotnet/machine-learning-and-data-science/keijiro-MeetBarracuda.html

[royshil]

thanks for the suggestion @iBenMoore

my intuition tells me that if we had a good segmentation model with very small input e.g. 32x32 or 48x48 - we can still reach a good visual experience (with contour smoothing) with very high frame rates. This is where i'll be spending my efforts in the short term.

[fangfufu]

Google's new Mediapipe library can achieve very high performance without GPU.
https://google.github.io/mediapipe/solutions/selfie_segmentation

I have replaced the backend of my webcam background replacement tool with it:
https://github.com/fangfufu/Linux-Fake-Background-Webcam

[royshil]

@fangfufu
Unfortunately

2021-06-08 21:00:00,604 - ERROR - Tensorflow op [segment: TFL_Convolution2DTransposeBias] is not supported
2021-06-08 21:00:00,604 - ERROR - Unsupported ops: Counter({'TFL_HARD_SWISH': 11, 'TFL_Convolution2DTransposeBias': 1})

[fangfufu]

@royshil , where do these error messages come from?

[royshil]

When i try to convert the .tflite model to .onnx with the official converter https://github.com/onnx/tensorflow-onnx

[fangfufu]

Ah ok.

[royshil]

Please see the latest release for the Meet model and another lightweight model.
The difference in performance is not very big, but the new models are faster.