【GroundTruth】
【My Prediction】
Its logic and details are written my blog.
https://trafalbad.hatenadiary.jp/entry/2020/02/23/094850
FPA
・Image size are more than 256
・Amount of calculation increase
HyperColumn
If you use multi input, using convolution layers output which are contrast position in model often get better practice.
Hypercolumn about better practice should be refered to Convolutional hypercolumns in Python