What is the best lite model for background removal?

[fatiherdogan01]

Hi,

I'm using "selfie_multiclass_256x256" and "deeplab_v3" models for background removal.
The "selfie multiclass 256x256" model works very well. But the "deeplabv3" model does not work well.

Does anyone have any suggestions for a lite model for mobile devices?

[edamlmmv]

Doesn't selfie_multiclass_256x256 work on mobile as well?
Personally I do not use deeplabv3, I preload MLKIT's model and once selfie_multiclass_256x256 is loaded I replace it.
(See volcomix's implementation for mlkit's model: https://github.com/Volcomix/virtual-background)

I think it's equivalent to SelfieSegmenter (square)

You may have to use webgl2 instead of canvas2d for the bigger model to be efficient on mobile