Research map / Detection and segmentation / CNN features
YOLOv10: Real-Time End-to-End Object Detection
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 YOLOv10, a real-time end-to-end YOLO detector that removes NMS via consistent dual label assignments (one-to-many and one-to-one heads) and uses holistic efficiency-accuracy driven architecture design.
Problem
YOLO detectors rely on non-maximum suppression (NMS) post-processing, which slows inference and hurts end-to-end deployment, and their architectures contain computational redundancy and limited capacity.
Approach
Train with a one-to-many head for rich supervision and a parallel one-to-one head used alone at inference (NMS-free), align both heads with a consistent matching metric (alpha_o2o=alpha_o2m, beta_o2o=beta_o2m) to reduce the supervision gap, and redesign YOLOv8 components (lightweight cls head, decoupled downsampling, rank-guided CIB blocks, large-kernel conv, partial self-attention) for efficiency
Key results
- YOLOv10-S reaches 46.3% AP at 2.49 ms, about 1.8x faster than RT-DETR-R18 at similar AP, with 2.8x fewer parameters and FLOPs.
- YOLOv10-B has 46% lower latency than YOLOv9-C at the same 52.5% AP with 25% fewer parameters.
- YOLOv10-L/X outperform YOLOv8-L/X by 0.3 and 0.5 AP with 1.8x and 2.3x fewer parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace YOLOv8-S with YOLOv10-S | YOLOv8-S · COCO val | AP: 44.9 -> 46.3 | Helped |
| use NMS-free training with dual assignments (ablation) | YOLOv10-S · COCO val | latency: 7.07 -> 2.44 | Helped |
| use NMS-free training with dual assignments (ablation) | YOLOv10-S · COCO val | AP: 44.9 -> 44.3 | Hurt |
| add efficiency-driven model design (ablation) | YOLOv10-S · COCO val | AP: 44.3 -> 44.5 | Helped |
| add accuracy-driven model design (ablation) | YOLOv10-S · COCO val | AP: 44.5 -> 46.3 | Helped |
| add one-to-one assignment to one-to-many training (ablation) | YOLOv8-S · COCO val | AP: 44.9 -> 44.3 | Hurt |
| use the consistent matching metric (α=0.5, β=6.0) rather than α=0.5, β=2.0 (ablation) | YOLOv10-S · COCO val | AP: 42.7 -> 44.3 | Helped |
| add a lightweight classification head (ablation) | YOLOv10-S · COCO val | AP: 44.3 -> 44.2 | No clear effect |
| add spatial-channel decoupled downsampling (ablation) | YOLOv10-S · COCO val | AP: 43.7 -> 44.4 | Helped |
| replace the baseline block with CIB (ablation) | YOLOv10-S · COCO val | AP: IRB 43.7; IRB-DW 44.2; ours 44.5 | Helped |
| add large-kernel convolution (ablation) | YOLOv10-S after efficiency design · COCO val | AP: 44.5 -> 44.9 | Helped |
| remove the reparameterization branch during training (ablation) | YOLOv10-S with large-kernel convolution · COCO val | AP: 0.1% AP degradation | Hurt |
Surprising findings
- NMS-free training with the one-to-one head alone causes a 1.0% AP gap versus one-to-many training for YOLOv10-N, shrinking to none for YOLOv10-X.
- Training YOLOv10-M with only the one-to-one head for longer (550 epochs) still yields inferior AP (48.9% vs 51.1%), indicating the one-to-many head's supervision matters.
Limitations
- Performance gap versus one-to-many NMS training remains, notably for small models (1.0% AP for N, 0.5% AP for S).
- Pretraining on large-scale datasets such as Objects365 was not explored due to limited compute.
- Latency is measured on specific hardware (T4 TensorRT) and training used 8 NVIDIA 3090 GPUs.
Experimental setup
Datasets: COCO
Benchmarks: COCO val2017
Metrics: AP AP50 AP75 APs/APm/APl latency (ms) parameters (M) FLOPs (G)
Baselines: YOLOv8 YOLOv9 YOLOv6-3.0 Gold-YOLO YOLO-MS RT-DETR DAMO-YOLO DEYO
Builds on: YOLOv8 TAL (task-aligned learning) DETR RT-DETR
Closest papers by meaning
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · similarity 0.90)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.90)
- YOLOv12: Attention-Centric Real-Time Object Detectors (NeurIPS 2025 · similarity 0.89)
- YOLO-RD: Introducing Relevant and Compact Explicit Knowledge to YOLO by Retriever-Dictionary (ICLR 2025 · similarity 0.88)
- YOLO-World: Real-Time Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.88)
- Roboflow100-VL: A Multi-Domain Object Detection Benchmark for Vision-Language Models (NeurIPS 2025 · similarity 0.88)
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.