In recent times, Detection Transformer(DETR) has been introduced and has shown excellent performance in object detection. As a result, there has been an increase in DETR-based object detectors such as AdaMixer, Deformable-DETR, and Sparse-RCNN. However, these detectors have complex architectures that lead to long inference times and suboptimal performance in detecting small objects. Previous approaches to this problem involve using dense priors for initialization and single decoder layer detectors, but their detection accuracy is low. To tackle this issue, we propose a novel approach called Trans-stage Interaction for object detection. This approach employs mixed supervision to improve the training efficiency and modeling capability of the model. We also introduce a trans-stage label assigner to improve the process of assigning labels accurately. With this label assigner, training target classes are gathered across stages and then reassigned to ensure accurate predictions in the subsequent stages. We also use a combination of one-to-one and one-to-many supervision to provide clear supervision during candidate generation, thereby enhancing the training efficiency of the model. Our model outperforms existing relevant models, as demonstrated through experimental results. Our model achieves 42.4 AP on the MS COCO benchmark with ResNet-50 as the backbone, 100 queries, and 12 epochs, improving the baseline by 1.2 AP. The experimental results verify our model over-performing the state-of-the-art related models.