Predictive Distillation Method of Anchor-Free Object Detection Model for Continual Learning

Continual learning (CL) is becoming increasingly important, not only for storage space because of the ever-increasing amount data being generated, but also associated copyright problems. In this study, we propose ground truth’ (GT’), which a combination truth (GT) and prediction teacher model that distills results previously trained model, called by applying knowledge distillation (KD) technique to an anchor-free object detection model. Among all objects predicted score higher than threshold value distilled into current student To avoid interference with new class learning, IoU obtained between every GT objects. Through continual scenario, even if reuse past limited, are sufficient, proposed minimizes catastrophic forgetting problems enables newly added classes. The was learned in PascalVOC 2007 + 2012 tested PascalVOC2007, better 9.6% p mAP 13.7% F1i shown scenario 19 1. result 15 5 showed compared algorithm, 1.6% 0.9% F1i. 10 outperformed other alternatives, 0.6%