Tensorflow attempt to reazlie ADVENT training for test dataset

ADVENT is Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation

Original paper is https://arxiv.org/abs/1811.12833

Main idea is to train a segmentation model with synthetic (source) data, but regularize it with loss of discriminator model, that will use entropy (uncertainty of models prediction) to help segmentation model do better with real (target) data.

Examples of models prediction

Unet-like model trained without adverasrial training

Unet-like model trained with adverasrial training