Leveraging uncertainty in adversarial learning to improve deep learning based segmentation

This paper proposes a new framework that combines Bayesian SegNet with adversarial learning to obtain high quality segmented objects of interest. The proposed architecture takes in the form of two discriminator networks that are trained separately. The first network discriminates between segmentation maps coming either from the SegNet or the ground truth. The second network discriminates between the model uncertainty obtained from SegNet and an ideal solution that does not include uncertainty. The process is very similar to the fusion of sensor information for better decision making. Uncertainty is considered as a measure of mistakes. Hence, learning from it will help improve the performance of neural networks. Our results show that we obtain higher accuracies compared to Bayesian SegNet. Training is performed on a small-sized dataset called CamVid and a large-sized dataset Sun RGB-D. The paper shows that dealing with uncertainties is beneficial for decision making in neural networks, especially in applications with highly uncertain environments. Examples include self-driving cars and medical imaging in cancer treatment.