Viet-Anh on Software Logo

What is: Conditional Instance Normalization?

SourceA Learned Representation For Artistic Style
Year2000
Data SourceCC BY-SA - https://paperswithcode.com

Conditional Instance Normalization is a normalization technique where all convolutional weights of a style transfer network are shared across many styles. The goal of the procedure is transform a layer’s activations xx into a normalized activation zz specific to painting style ss. Building off instance normalization, we augment the γ\gamma and β\beta parameters so that they’re N×CN \times C matrices, where NN is the number of styles being modeled and CC is the number of output feature maps. Conditioning on a style is achieved as follows:

z=γ_s(xμσ)+β_s z = \gamma\_{s}\left(\frac{x - \mu}{\sigma}\right) + \beta\_{s}

where μ\mu and σ\sigma are xx’s mean and standard deviation taken across spatial axes and γ_s\gamma\_{s} and β_s\beta\_{s} are obtained by selecting the row corresponding to ss in the γ\gamma and β\beta matrices. One added benefit of this approach is that one can stylize a single image into NN painting styles with a single feed forward pass of the network with a batch size of NN.