Fig. 1.

Diagram of proposed method. At training time, x_u, x_l and y_l are supplied to the network. x_u is an image from the unlabeled target domain and x̂_u is the result of applying some augmentation function to x_u. A labeled image, x_l, is passed through the network, f_θ before combining with a label y_l to form the segmentation loss, ℒ_s. The image representations are fed to a domain discriminator d _Ω which attempts to maximise the cross-entropy between predicted domain and actual domain, ℒ_adv. Finally, similarity is promoted between the network predictions on x_u and ${\widehat{x}}_u$ using ℒ_PC.