We present DINO (\textbf{D}ETR with \textbf{I}mproved de\textbf{N}oising anch\textbf{O}r boxes), a state-of-the-art end-to-end object detector. % in this paper. DINO improves over previous DETR-like models in performance and efficiency by using a contrastive way for denoising training, a mixed query selection method for anchor initialization, and a look forward twice scheme for box prediction. DINO achieves $49.4$AP in $12$ epochs and $51.3$AP in $24$ epochs on COCO with a ResNet-50 backbone and multi-scale features, yielding a significant improvement of $\textbf{+6.0}$\textbf{AP} and $\textbf{+2.7}$\textbf{AP}, respectively, compared to DN-DETR, the previous best DETR-like model. DINO scales well in both model size and data size. Without bells and whistles, after pre-training on the Objects365 dataset with a SwinL backbone, DINO obtains the best results on both COCO \texttt{val2017} ($\textbf{63.2}$\textbf{AP}) and \texttt{test-dev} (\textbf{$\textbf{63.3}$AP}). Compared to other models on the leaderboard, DINO significantly reduces its model size and pre-training data size while achieving better results. Our code will be available at \url{https://github.com/IDEACVR/DINO}.
核心贡献 · Key contributions
提出对比去噪训练，通过正负样本改善一对一匹配。 Proposes contrastive denoising training with positive and negative samples to improve one-to-one matching.
引入混合查询选择，从编码器特征初始化锚框，同时保持内容查询可学习。 Introduces mixed query selection to initialize anchor boxes from encoder features while keeping content queries learnable.
开发前视两次方案，利用后层梯度更新前层参数。 Develops look forward twice scheme to leverage later layer gradients for earlier layer parameter updates.
在 COCO 上用 ResNet-50 在 12 轮达到 49.4 AP，24 轮达到 51.3 AP，优于先前类 DETR 模型。 Achieves 49.4 AP in 12 epochs and 51.3 AP in 24 epochs on COCO with ResNet-50, outperforming prior DETR-like models.
扩展到大型骨干网络和数据，使用 SwinL 和 Objects365 在 COCO val2017 上达到 63.2 AP，test-dev 上达到 63.3 AP。 Scales to large backbones and data, achieving 63.2 AP on COCO val2017 and 63.3 AP on test-dev with SwinL and Objects365.
首个在 COCO 排行榜上超越 SOTA 的端到端 Transformer 检测器，且模型和数据规模更小。 First end-to-end Transformer detector to surpass SOTA on COCO leaderboard with reduced model and data size.
局限 · Limitations
评估仅限于 COCO 数据集，未验证对其他领域的泛化性。 Evaluation limited to COCO dataset; generalization to other domains not verified.
性能依赖于 Objects365 预训练，无大规模预训练时可能不成立。 Performance depends on pre-training on Objects365; may not hold without large-scale pre-training.
对比去噪需要调整噪声尺度λ1 和λ2，敏感性未充分探索。 Contrastive denoising requires tuning noise scales λ1 and λ2; sensitivity not fully explored.
前视两次因额外梯度流增加计算开销。 Look forward twice adds computational overhead due to extra gradient flow.
混合查询选择假设编码器特征提供可靠空间先验，编码器差时可能失效。 Mixed query selection assumes encoder features provide reliable spatial priors; may fail with poor encoder.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2.1 经典目标检测器2.1 Classical Object Detectors
2.2 DETR 及其变体2.2 DETR and Its Variants
2.3 大规模预训练用于目标检测2.3 Large-scale Pre-training for Object Detection
3 DINO：改进去噪锚框的 DETR3 DINO: DETR with Improved DeNoising Anchor Boxes