In recent years, object detection is transferred to other fields, for example, face and vehicle detection. However, the bounding-box labeling is a huge resources cost work. This study solves the problem that transfer object detection task to other domain dataset without bounding-box label. A relationship layer is built to learn the relationship between classification and regression task. In addition, we construct a product dataset, on which rotatable object detection is solved using our training method. A proposal selecting method is proposed for training classification based on faster RCNN framework without bounding-box label. We propose a object detection method without bounding-box annotation. The method is easy to transfer to other datasets and training.