search

lizhengwei1992 / Semantic_Human_Matting

Semantic Human Matting
537 stars 146 forks source link

M-net pre-train #15

Open ulnay opened 5 years ago

ulnay commented 5 years ago

First of all, thank you very much for implementation and sharing. I just got one question, in the paper, section 4.5 , they've mentioned they also have M-net pretrain, in your implementation, how can we pre-train the M-net? Thank you!

Jason-xin commented 5 years ago

the same question

rainsun1 commented 5 years ago
  1. T-Net pre-train: the model shoule only include the t-net part, and you can only train the model with L_t loss.
  2. M-Net pretrain: the model shoule only include the m-net part, and the input must include 6 mixed channels including 3-channel image and 3-channel trimap(fg, bg, unsure), and feed them to m-net. L_alpha and L_composition losses are included here.
  3. Fine tune: include both parts, with the combined two weights above.