关键词: 标签选择, 目标检测, 伪标签, 灾难性遗忘, 增量学习, 注意力模块

Abstract: Incremental learning has achieved good results in image classification,but it is challenging to apply incremental learning to multi-class object detection.Object detection is more complex than image classification,which combines classification and border regression.At present,the most advanced incremental object detectors adopt the external fixed region suggestion method based on knowledge distillation,which consumes a lot of time and cost.For single-stage detectors,due to the lack of annotation and region advice information for the old class,old objects are usually identified by the detector as the background,resulting in catastrophic forgetting.In this paper,a label selection algorithm based on border distance metric is proposed.It uses the detection results of the old model and the existing dataset labels to select and merge by measuring the coincidence of the bounding boxes,making up for the lack of annotations of the old objects in the new dataset and alleviating catastrophic forgetting.In addition,a module that combines the attention module with the residual module is designed to extract discriminative features at different depths in feature extraction network,to further improve the detection accuracy of model.The proposed method is implemented in the single-stage detection framework,and the effectiveness of the method is verified on PASCAL VOC dataset.Compared with the best model at present,the average accuracy value of the old object and all objects improves by 2.8% and 2.1%,respectively.The pseudo-labels obtained by the proposed method greatly alleviate the forgetting problem,and the attention residual module improves the detection accuracy of the model.

Key words: Attention module, Catastrophic forgetting, Incremental learning, Label selection, Object detection, Pseudo label