Continual Learning for Instance Segmentation to Mitigate Catastrophic Forgetting
Jeong Jun Lee, Seung Il Lee, Hyun Kim · 2021
In recent years, with the development of GPUs, there are various deep learning models that show excellent performance in the field of computer vision, but these models have a fatal limitation of catastrophic forgetting. Continual learning studies for mitigating the catastrophic forgetting have been conducted on classification tasks, but in real-life, object detection and segmentation tasks are more practical than classification tasks. This paper proposes a catastrophic forgetting solution in the instance segmentation network by extending and applying the catastrophic forgetting solution, which was previously limited to classification, to segmentation tasks.