Learning-Without-Forgetting via Memory Index in Incremental Object Detection

Object detection has made significant progress in recent years. However, when the training data is continuous and dynamic, notorious catastrophic forgetting will occur. In addition, training together with old class data requires more storage space and training time, and even the old data is not available. To address the above challenges in incremental object detection (IOD), this paper proposes an one-stage and anchor-based IOD paradigm via parameter isolation with memory index (LwF-MI) that does not require any old data. The core of LwF-MI is to design the anchor-based classification masks for memory indexing to guide the incremental network output. Besides, we propose a multi-scale fusion method for the preprocessing of the above masks, which addresses the poor performance of large-scale masks. Extensive experiments on MS COCO demonstrate that our approach achieves state-of-the-art results, which even exceed the full training results.