Abstract:
Deep neural networks (DNN) in object detecting tasks have witnessed significant progress in the past years. However, it relies on intensive training data with accurate bounding box annotations for a remarkable performance. Once the labelled data are hard to catch, the generalization ability of DNN is far from satisfactory. We propose a few-shot object detecting method based on a multi-stage training strategy within feature redistribution (MSFR). Based on the analyses of the distribution of source domain dataset and target domain dataset in few-shot tasks, a feature redistribution algorithm is proposed to make the feature distribution meet Gaussian distribution or quasi-Gaussian distribution. It solves the inconsistency distribution of the source domain dataset and the target domain dataset. Then, a multi-stage training algorithm is proposed, which improves the efficiency of transferring the source domain knowledge to the target domain task when only a small amount of labeled data for training in each class. Thus, our proposed method significantly improves the detection performance of few-shot target domain categories while maximizing the detection accuracy of the source domain categories. The experimental results on VOC datasets show that the proposed algorithm achieves a precision improvement of up to 9.06% on different tasks, compared with existing few-shot object detection approaches.