Decoupled Mutual Distillation for Incremental Object Detection

Visual object detection aims to localize and classify objects in the images. Popular detectors perform well when all the object categories are pre-defined. However, they show poor performance when the new categories join in the training incrementally and the replay of old training samples is not allowed. This issue is widely known as catastrophic forgetting. In this paper, the detection head of classic Faster R-CNN is decoupled into the classification head and localization head. Based on the modified network, a dynamic mutual distillation framework is designed. The distillation is applied only to the classification head. On the one hand, this design allows the knowledge to be transferred from both the old model and the assistant model to the new. On the other hand, the degree of knowledge-transfer from either the old or assistant model is dynamically determined by their expertise on a certain category. Compared to the conventional teacher-student distillation, our design makes a better balance between the inheritance of knowledge from the old model and adapting to the new categories.