Abstract
We introduce D-FINE, a powerful real-time object detector that achievesoutstanding localization precision by redefining the bounding box regressiontask in DETR models. D-FINE comprises two key components: Fine-grainedDistribution Refinement (FDR) and Global Optimal Localization Self-Distillation(GO-LSD). FDR transforms the regression process from predicting fixedcoordinates to iteratively refining probability distributions, providing afine-grained intermediate representation that significantly enhanceslocalization accuracy. GO-LSD is a bidirectional optimization strategy thattransfers localization knowledge from refined distributions to shallower layersthrough self-distillation, while also simplifying the residual prediction tasksfor deeper layers. Additionally, D-FINE incorporates lightweight optimizationsin computationally intensive modules and operations, achieving a better balancebetween speed and accuracy. Specifically, D-FINE-L / X achieves 54.0% / 55.8%AP on the COCO dataset at 124 / 78 FPS on an NVIDIA T4 GPU. When pretrained onObjects365, D-FINE-L / X attains 57.1% / 59.3% AP, surpassing all existingreal-time detectors. Furthermore, our method significantly enhances theperformance of a wide range of DETR models by up to 5.3% AP with negligibleextra parameters and training costs. Our code and pretrained models:https://github.com/Peterande/D-FINE.