Most ML models process frames as independent images
This one uses recurrent neural network to process videos with temporal memory
As a result, its far more stable between frames and also achieves higher performance

• Credit: https://github.com/PeterL1n/RobustVideoMatting

• model/mb3-i8 actual rvm model converted to tfjs graph model and quantized to uint8
• src/rvm.ts main module that exposes two methods, load and predict
  module behavior is controlled via configuration params passed to predict method
• src/index.ts demo app that uses rvm
  actual canvas merging (overlays, etc.) is done using native canvas globalCompositeOperation
• src/webcam.ts helper class used by demo that deals with webcam initialization and control