Object detection is an important computer vision task which is required in various applications such as autonomous vehicle, robotics, etc. In this paper, a data-efficient object detection methodology with few training samples is proposed to achieve the data-efficient object detection. On this matter, it is commonly known and encountered that the performance of conventional fully supervised approaches typically suffer when data is insufficient. Classification confidence and localization accuracy are expectedly not consistent with unseen categories, which often leads to overfitting of class-specific parameters. In this paper, we thus propose a novel data-efficient classification network using weight imprinting which can be adopted into any anchor-based detector to directly enhance the detection accuracy on data-rare categories, without sacrificing the performance on base categories. The proposed network is data-efficient yet highly effective with the carefully designed weight imprinting head. Experiments show our proposed method can bring significant performance gains to rarely labeled categories on COCO and PASCAL VOC.
