Fig. 3.

Architecture of DEP-GAN, which consists of one generator (upper side, “A”) and two critics (lower side, “C” and “D”). Note how the proposed auxiliary input is feed-forwarded to convolutional layers (yellow, “B”) and then modulated to the generator using FiLM layer (green) inside residual block (ResBlock) (light blue, “E”). Please see Fig. 2 for connections between each part and Section 2.3 for full explanation about auxiliary input. On the other hand, DEP-UResNet is based on DEP-GAN’s generator, including its auxiliary input, with modification of the last non-linear activation function (i.e., use softmax for segmentation instead of tanh). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)