Research map / Detection and segmentation / CNN features
D-FINE: Redefine Regression Task of DETRs as Fine-grained Distribution Refinement
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes D-FINE, a real-time DETR detector that reframes box regression as iterative residual refinement of per-edge probability distributions (FDR) and distills localization from the final decoder layer into shallower layers (GO-LSD).
Problem
Bounding box regression in DETR-style real-time detectors uses fixed coordinates (Dirac delta), which cannot model localization uncertainty and converges slowly; distillation for real-time DETRs is also costly or ineffective.
Approach
Model each box edge as a non-uniform discrete distribution refined residually across decoder layers (FDR), and use the final layer's refined distributions as soft targets for self-distillation into shallower layers with IoU- and confidence-decoupled weights and a union of Hungarian matches (GO-LSD).
Key results
- D-FINE-L reaches 54.0% AP on COCO val2017 at 8.07 ms (124 FPS) on T4 with TensorRT FP16, 31M params and 91 GFLOPs.
- With Objects365 pretraining, D-FINE-X reaches 59.3% AP and D-FINE-L 57.1% AP, outperforming YOLOv10-X (54.9%) and YOLOv10-L (54.0%) at the same pretraining setting.
- Adding FDR and GO-LSD to DETR models improves AP by 2.0 to 5.3 points (e.g. DAB-DETR 44.2 to 49.5) with no added parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FDR and GO-LSD | Deformable-DETR · COCO val2017 | AP: 43.7 -> 47.1 (+3.4) | Helped |
| add FDR and GO-LSD | DAB-DETR · COCO val2017 | AP: 44.2 -> 49.5 (+5.3) | Helped |
| add FDR and GO-LSD | DN-DETR · COCO val2017 | AP: 46.0 -> 49.7 (+3.7) | Helped |
| add FDR and GO-LSD | DINO · COCO val2017 | AP: 49.0 -> 51.6 (+2.6) | Helped |
| remove decoder projection layers (ablation) | RT-DETR-HGNetv2-L baseline · COCO val2017 | AP: 53.0 -> 52.4 | Hurt |
| add Target Gating Layers (ablation) | RT-DETR-HGNetv2-L without decoder projection layers · COCO val2017 | AP: 52.4 -> 52.8 | Helped |
| replace encoder CSP layers with GELAN (ablation) | RT-DETR-HGNetv2-L variant · COCO val2017 | AP: 52.8 -> 53.5 | Helped |
| halve GELAN hidden dimension (ablation) | GELAN encoder variant · COCO val2017 | AP: 53.5 -> 52.8 | Hurt |
| use uneven sampling points (S: 3, M: 6, L: 3) (ablation) | Reduced-dimension GELAN · COCO val2017 | AP: 52.8 -> 52.9 | Helped |
| use alternative sampling (S: 6, M: 3, L: 3) or (S: 3, M: 3, L: 6) (ablation) | D-FINE sampling configuration · COCO val2017 | AP: -0.1% AP | Hurt |
| adopt RT-DETRv2 training strategy (ablation) | D-FINE sampling configuration · COCO val2017 | AP: 52.9 -> 53.0 | Helped |
| add FDR (ablation) | RT-DETRv2 Training Strategy variant · COCO val2017 | AP: 53.0 -> 53.5 | Helped |
Surprising findings
- Traditional logit mimicking and feature imitation distillation reduce or fail to improve AP for DETR models (Logit Mimicking drops AP to 52.6%), attributed to unstable one-to-one matching.
Limitations
- Performance gap between lighter D-FINE models and other compact detectors remains small.
- Shallow decoder layers may yield less accurate final-layer predictions, limiting distillation into early layers.
Experimental setup
Datasets: COCO train2017 COCO val2017 Objects365
Benchmarks: COCO val2017 Objects365
Metrics: AP AP50 AP75 APS APM APL latency (ms) GFLOPs parameters
Baselines: YOLOv10 RT-DETR RT-DETRv2 LW-DETR YOLO11 YOLOv8 YOLOv9 Gold-YOLO RTMDet Logit Mimicking Feature Imitation Localization Distillation
Builds on: RT-DETR GFocal Distribution Focal Loss Localization Distillation Deformable DETR
Closest papers by meaning
- DiffuBox: Refining 3D Object Detection with Point Diffusion (NeurIPS 2024 · similarity 0.85)
- From Coarse to Fine: Robust Hierarchical Localization at Large Scale (CVPR 2019 · similarity 0.85)
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.84)
- Benchmarking Object Detectors under Real-World Distribution Shifts in Satellite Imagery (CVPR 2025 · similarity 0.83)
- Stochastic Localization via Iterative Posterior Sampling (ICML 2024 · similarity 0.83)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.