ECSSGAN is a deep convolutional neural network for visual saliency prediction which is trained on adversarial examples. It is based on SALGAN architecture and is able to predict the visual saliency of images with extended complex scenes. For the purposes of the network, ECSS dataset was used and the end network was able to predict the saliencies of the images with depth information(portrait mode photos).
The first stage of the network consists of a generator model with an architecture of an autoencoder based on VGG16. It generates saliency maps when fed with images of complex scenes. The resulting prediction is processed by a discriminator network which is a CNN, trained to perform binary classification task between the saliency maps generated by the generator and to that of the ground truth labels.