Well actually, just clone leekunhee/Mask_RCNN repo and even without applying the minor changes we did you can grab the keras model by doing model.keras_model.

And then you have to pass that to TFLiteConverter.

One thing important is to manually (could be calculated automatically) set the input sizes. You'll need to edit mrcnn/model.py for this.

I will try to provide a colab notebook.