megvii-research/RealFlow

About M-step

Eryo-iPython opened this issue · 0 comments

Eryo-iPython commented

thanks for your good work, I have some questions for M-step

  1. Using C+T to generate new dataset G1.
  2. Using G1 to train RAFT get new weight C+T+G1.
  3. Using C+T+G1 to generate dataset G2.
    my question is that which weight should used to train RAFT in the next step? In the other words, I should use the weight C+T+G1 to train RAFT or keep C+T to train it on dataset G2？