Research map / Detection and segmentation / CNN features
YOLOv12: Attention-Centric Real-Time Object Detectors
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 YOLOv12, an attention-centric real-time object detector with area attention, R-ELAN aggregation, and YOLO-specific architectural changes, trained from scratch on COCO.
Problem
Attention-based detectors are too slow for real-time YOLO systems due to quadratic complexity and inefficient memory access, so YOLO architecture work has remained CNN-centric.
Approach
Area attention splits the feature map into l=4 horizontal or vertical strips, avoiding explicit window partitioning and cutting attention cost via simple reshapes; combined with residual ELAN (R-ELAN) with scaled residuals and bottleneck-style aggregation, FlashAttention, removal of positional encodings, MLP ratio 1.2, and a 7x7 separable-conv position perceiver.
Key results
- YOLOv12-N reaches 40.6% mAP at 1.64 ms on T4 with TensorRT FP16, 2.1 mAP above YOLOv10-N and 1.2 above YOLOv11-N.
- YOLOv12-S reaches 48.0% mAP, beating RT-DETR-R18 by 1.5 mAP while running 42% faster with 36% of FLOPs and 45% of parameters.
- Area attention speeds up YOLOv12-N by about 0.7 ms on RTX 3080 FP32 with only a 0.2 mAP drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace YOLO11 with YOLOv12 | YOLO detector family · COCO val | AP50:95: 39.4 -> 40.6; 46.9 -> 48.0; 51.5 -> 52.5; 53.3 -> 53.7; 54.6 -> 55.2 | Helped |
| add area attention (ablation) | YOLOv12-N/S/X · CUDA latency, FP32 | latency (ms): 2.7/2.5 -> 2.0/2.0; 5.1/4.4 -> 3.5/3.1; 26.4/21.9 -> 18.2/14.3 | Helped |
| replace vanilla ELAN with Re-Aggre. (ablation) | YOLOv12-N · COCO val | AP50:95: 40.8 -> 40.6 | Hurt |
| add residual block with scaling 0.1 (ablation) | YOLOv12-N · COCO val | AP50:95: 40.6 -> 40.3 | Hurt |
| use residual block with scaling 0.01 in R-ELAN (ablation) | YOLOv12-L/X · training convergence | convergence | Helped |
| use Conv+BN instead of Linear+LN attention (ablation) | YOLOv12-N · COCO val | AP50:95: 40.5 -> 40.6 | Helped |
| remove hierarchical design (ablation) | YOLOv12-N · COCO val | AP50:95: 40.6 -> 38.3 | Hurt |
| omit stage S1 (ablation) | YOLOv12-N · COCO val | AP50:95: 40.6 -> 40.1 | Hurt |
| omit stage S4 (ablation) | YOLOv12-N · COCO val | AP50:95: 40.6 -> 39.8 | Hurt |
| train for 600 rather than 300 or 500 epochs (ablation) | YOLOv12-N/S · COCO val | AP: 39.5 -> 40.6; 47.0 -> 48.0 | Helped |
| increase Position Perceiver kernel from 3x3 to 7x7 (ablation) | YOLOv12-N · COCO val | AP50:95: 40.4 -> 40.6 | Helped |
| add positional embedding (ablation) | YOLOv12-N · COCO val | AP50:95: 40.6 -> 40.3 (RPE); 40.6 -> 40.5 (APE) | Hurt |
Surprising findings
- Plain (non-hierarchical) ViT-style design drops YOLOv12 to 38.3% mAP, unlike Mask R-CNN where plain ViTs work well.
- Layer norm performs worse than batch norm when attention is combined with convolution in this YOLO setting.
Limitations
- Requires FlashAttention, which only supports Turing, Ampere, Ada Lovelace, or Hopper GPUs.
- Area attention still has quadratic cost in n, which grows with input resolution.
- L- and X-scale models failed to converge or were unstable without R-ELAN residual scaling.
Experimental setup
Datasets: MS COCO 2017
Benchmarks: MS COCO 2017 val
Metrics: AP 50:95 AP50 AP75 AP small/medium/large latency (ms) FLOPs parameters
Baselines: YOLOv6-3.0 Gold-YOLO YOLOv8 YOLOv9 YOLOv10 YOLOv11 RT-DETR RT-DETRv2
Builds on: YOLOv11 ELAN FlashAttention layer scaling C3K2 / GELAN
Closest papers by meaning
- YOLOv10: Real-Time End-to-End Object Detection (NeurIPS 2024 · similarity 0.89)
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · similarity 0.87)
- DETRs Beat YOLOs on Real-time Object Detection (CVPR 2024 · similarity 0.86)
- YOLO-RD: Introducing Relevant and Compact Explicit Knowledge to YOLO by Retriever-Dictionary (ICLR 2025 · similarity 0.84)
- YOLO-World: Real-Time Open-Vocabulary Object Detection (CVPR 2024 · similarity 0.84)
- Samba: A Unified Mamba-based Framework for General Salient Object Detection (CVPR 2025 · similarity 0.82)
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.