Abstract: There are two main problems with current distillation methods for object detection: first, most focus on the distillation of the feature extraction part; second, most rely on ground truth (GT) to determine the distillation region. To address the above problems, we proposed confidence ranking distillation, and created an adaptive distillation model based on ranking for the proposal boxes generated by the region proposal network (RPN); proposed confidence-guided feature distillation, exploring a new paradigm of using the proposal boxes generated by the teacher network to guide distillation regions. Experiments show that the algorithm improves by 8.5 percentage points on the PASCAL VOC dataset compared with the baseline model.