this picture may be different with others, you should focus this it add
2x conv7 not
2x pool5 in
FCN-16s. Actually the
conv7's output_shape is
(w/32,h/32,classes_number) in my implementation. I know
conv7 should multiply 2 to achieve the same shape (1st,2nd) with
pool4. But my question is the
pool4's shape is
(w/16,h/16,512), ps 512 is the output_channel of the last conv-layer.
2x conv7 have not same shape(3rd), how to add?
And another question is what are the train-labels for the FCN, I think a train-label is just a image which has the same width and same height with input_image and has classes_number channel to show each pixel's possibility belong to each class. This idea is derived from the conv7's output_shape, is it right? I hope you can give me more and clear explanations. Thanks!!!