Rotation-fused Consistency Semi-supervised Learning for Object Detection

In recent years, the neural network has been widely used in the object detection field. Most methods are supervised learning which need lots of labeled data. However, the production of the labeled data set is a costly work, especially in the object detection task. The utilization of unlabeled data has attracted much attention from the academic and industrial areas. Semi-supervised learning is one of the effective ways to simultaneously use both labeled and unlabeled data in the training process which has achieved great success in the field of image classification. Unfortunately, these methods used in image classification are not suitable for object detection completely. In this paper, we propose the Rotation-fused Consistency Semi-supervised Learning for Object Detection(RCSD), which uses rotation fusion, applies the consistency regularization to object detection, and takes into account both the classification and location of objects. The effectiveness of our method in the data set PASCAL VOC is verified. The results show that the performance of the detector is enhanced by using rotation consistency and it is further improved by multi-rotation fusion.