Causes of Catastrophic Forgetting in Class-Incremental Semantic Segmentation

Class-incremental learning for semantic segmentation (CiSS) is presently a highly researched field which aims at updating model by sequentially new classes. A major challenge in CiSS overcoming the effects of catastrophic forgetting, describes sudden drop accuracy on previously learned classes after trained set Despite latest advances mitigating underlying causes forgetting specifically are not well understood. Therefore, experiments and representational analyses, we demonstrate that shift background class bias towards CiSS. Furthermore, show both mostly manifest themselves deeper classification layers network, while early affected. Finally, how effectively mitigated utilizing information contained background, with help knowledge distillation an unbiased cross-entropy loss.