SID: Incremental learning for anchor-free object detection via Selective and Inter-related Distillation

Incremental learning requires a model to continually learn new tasks from streaming data. However, traditional fine-tuning of well-trained deep neural network on task will dramatically degrade performance the old — problem known as catastrophic forgetting. In this paper, we address issue in context anchor-free object detection, which is trend computer vision it simple, fast, and flexible. Simply adapting current incremental strategies fails these detectors due lack consideration their specific structures. To deal with challenges detectors, propose novel paradigm called Selective Inter-related Distillation (SID). addition, evaluation metric proposed better assess under conditions. By selective distilling at proper locations further transferring additional instance relation knowledge, our method demonstrates significant advantages benchmark datasets PASCAL VOC COCO. • We explore detection fully convolutional detectors. A inter-related distillation strategy proposed. evaluate results. demonstrate superior datasets.