This paper introduces Relation-DETR, a novel method for enhancing the performance and convergence of DETR (DEtection TRansformer) by incorporating explicit position relation prior. The key idea is to address the slow convergence issue in transformers by introducing position relation embeddings that provide structural bias to the self-attention mechanism. The method constructs position relation embeddings through an encoder, which is then used for progressive attention refinement. This approach extends the traditional streaming pipeline of DETR into a contrastive relation pipeline, enabling better handling of non-duplicate predictions and positive supervision. The proposed method, Relation-DETR, achieves significant improvements in detection performance on various datasets, including COCO val2017, with a 2.0% increase in AP compared to DINO. It also demonstrates faster convergence, achieving over 40% AP with only 2 training epochs. The position relation encoder is designed as a universal plug-in-and-play component, offering clear improvements for any DETR-like method. Additionally, the method is evaluated on a new class-agnostic detection dataset, SA-Det-100k, where it achieves a 1.3% AP improvement. The paper also explores the statistical significance of object position relations, proposing a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient. The results show that object position relations are statistically significant across various datasets, indicating their importance in object detection. The method is evaluated on multiple datasets, including COCO 2017 and task-specific datasets like CSD and MSSD. The results demonstrate that Relation-DETR outperforms existing DETR methods in terms of detection performance, convergence speed, and transferability. The position relation encoder is shown to be effective in improving detection performance for both generic and task-specific scenarios. The paper also discusses the transferability of the position relation encoder, showing that it can be integrated into various DETR-based methods with minimal modifications, leading to consistent performance improvements. The method is further validated through ablation studies and experiments on different datasets, demonstrating its effectiveness and generalizability. Overall, the paper presents a novel approach to enhancing DETR detectors by incorporating explicit position relation prior, leading to improved performance, faster convergence, and better generalizability. The method is shown to be effective across various datasets and scenarios, making it a promising advancement in the field of object detection.This paper introduces Relation-DETR, a novel method for enhancing the performance and convergence of DETR (DEtection TRansformer) by incorporating explicit position relation prior. The key idea is to address the slow convergence issue in transformers by introducing position relation embeddings that provide structural bias to the self-attention mechanism. The method constructs position relation embeddings through an encoder, which is then used for progressive attention refinement. This approach extends the traditional streaming pipeline of DETR into a contrastive relation pipeline, enabling better handling of non-duplicate predictions and positive supervision. The proposed method, Relation-DETR, achieves significant improvements in detection performance on various datasets, including COCO val2017, with a 2.0% increase in AP compared to DINO. It also demonstrates faster convergence, achieving over 40% AP with only 2 training epochs. The position relation encoder is designed as a universal plug-in-and-play component, offering clear improvements for any DETR-like method. Additionally, the method is evaluated on a new class-agnostic detection dataset, SA-Det-100k, where it achieves a 1.3% AP improvement. The paper also explores the statistical significance of object position relations, proposing a quantitative macroscopic correlation (MC) metric based on the Pearson Correlation Coefficient. The results show that object position relations are statistically significant across various datasets, indicating their importance in object detection. The method is evaluated on multiple datasets, including COCO 2017 and task-specific datasets like CSD and MSSD. The results demonstrate that Relation-DETR outperforms existing DETR methods in terms of detection performance, convergence speed, and transferability. The position relation encoder is shown to be effective in improving detection performance for both generic and task-specific scenarios. The paper also discusses the transferability of the position relation encoder, showing that it can be integrated into various DETR-based methods with minimal modifications, leading to consistent performance improvements. The method is further validated through ablation studies and experiments on different datasets, demonstrating its effectiveness and generalizability. Overall, the paper presents a novel approach to enhancing DETR detectors by incorporating explicit position relation prior, leading to improved performance, faster convergence, and better generalizability. The method is shown to be effective across various datasets and scenarios, making it a promising advancement in the field of object detection.