Knowledge distillation for incremental learning in semantic segmentation

Deep learning architectures have shown remarkable results in scene understanding problems, however they exhibit a critical drop of performances when are required to learn incrementally new tasks without forgetting old ones. This catastrophic phenomenon impacts on the deployment artificial intelligence real world scenarios where systems need and different representations over time. Current approaches for incremental deal only with image classification object detection tasks, while this work we formally introduce semantic segmentation. We tackle problem applying various knowledge distillation techniques previous model. In way, retain information about learned classes, whilst updating current model developed four main methodologies working both output layers internal feature representations. do not store any belonging training stages last is used preserve high accuracy previously classes. Extensive experimental Pascal VOC2012 MSRC-v2 datasets show effectiveness proposed several scenarios.