Causes of Catastrophic Forgetting in Class-Incremental Semantic Segmentation
This addresses the problem of catastrophic forgetting for researchers and practitioners in incremental learning for semantic segmentation, though it is incremental as it builds on existing mitigation techniques.
The paper investigates the causes of catastrophic forgetting in class-incremental semantic segmentation, identifying semantic shift of the background class and bias towards new classes as major factors, and shows that these issues can be mitigated using knowledge distillation and an unbiased loss, improving performance by up to 15% in mIoU on benchmark datasets.
Class-incremental learning for semantic segmentation (CiSS) is presently a highly researched field which aims at updating a semantic segmentation model by sequentially learning new semantic classes. A major challenge in CiSS is overcoming the effects of catastrophic forgetting, which describes the sudden drop of accuracy on previously learned classes after the model is trained on a new set of classes. Despite latest advances in mitigating catastrophic forgetting, the underlying causes of forgetting specifically in CiSS are not well understood. Therefore, in a set of experiments and representational analyses, we demonstrate that the semantic shift of the background class and a bias towards new classes are the major causes of forgetting in CiSS. Furthermore, we show that both causes mostly manifest themselves in deeper classification layers of the network, while the early layers of the model are not affected. Finally, we demonstrate how both causes are effectively mitigated utilizing the information contained in the background, with the help of knowledge distillation and an unbiased cross-entropy loss.