hyojinpark / ttvos Goto Github PK

The training set only have 3 frames, so the network have 2 frame to predict, so the state['prev_seg'] should only have 4 channels because the last predicted mask is not used to update the long-term template. Did I misunderstand your code? Why you use 6 channels state['prev_seg']?

Does this part (line 187 in models/TTVOS_model.py) is to update the long-term template? It looks like that it is different from what in you article. Do you use transition matrix (state['update_W']) to update the long-term template? Or I misunderstood your code, this part is not used to update the template, or state['update_W'] is not the transition matrix？