what is the initial model of the PSE model? the Views Predictor or Pose Map Models?

[datar001]

hi , I suddenly have a question. The above picture says that ' use an initial model with a trained views predictor' . But the input of pse model is 17 channels, not matching with 3 channels in views model . whether the 'conv1' should be added in checkpoint-exclude-scopes? or can we use the pose model as the initial model of pse model? thanks very much.