Our approach models label noise in semantic segmentation with a probabilistic framework that incorporates spatial correlations. By introducing a continuous latent ...