Evidential Deep Learning for Class-Incremental Semantic Segmentation

Class-Incremental Learning is a challenging problem in machine learning that aims to extend previously trained neural networks with new classes. This especially useful if the system able classify objects despite original training data being unavailable. Although semantic segmentation has received less attention than classification, it poses distinct problems and challenges, since previous future target classes can be unlabeled images of single increment. In this case, background, past are correlated there exists background-shift. paper, we address how model while avoiding spurious feature clustering uncorrelated We propose use Evidential Deep evidence as Dirichlet distribution. Our method factorizes into separate foreground class probability, calculated by expected value distribution, an unknown (background) probability corresponding uncertainty estimate. our novel formulation, background implicitly modeled, space comes from forcing output high score for pixels not labeled objects. Experiments on incremental Pascal VOC ADE20k benchmarks show superior state art, when repeatedly increasing number increments.