Instance Normalization

[fujistoo]

In the paper, you have mentioned that each convolutional block for the DHAN follows by an Instance Normalization (IN) and LeakyReLU layer, however, it uses Batch Normalization here instead.

ghost-free-shadow-removal/networks.py

Line 25 in 93cc1d6

return w0*x+w1*slim.batch_norm(x)

Am I missing out on something here or is this the representation for IN? I do know that BN is used for shadow removal, but nm was referred on the build_aggasatt_joint, so it confuses me a little.