Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

YOLOv12: Attention-Centric Real-Time Object Detectors

Yunjie Tian, Qixiang Ye, David Doermann

TL;DR

This work tackles the tension between speed and accuracy in real-time object detection by infusing YOLO with attention while preserving CNN-style efficiency. It introduces Area Attention (A2) to reduce self-attention cost, Residual Efficient Layer Aggregation Networks (R-ELAN) to stabilize optimization, and targeted architectural refinements (e.g., removing positional encodings, adjusting MLP ratio) to fit the YOLO pipeline. Through extensive COCO experiments across five model scales, YOLOv12 achieves state-of-the-art latency-accuracy trade-offs, outperforming existing real-time detectors and RT-DETR variants on multiple metrics and hardware configurations. The results demonstrate that attention-centric designs can meet stringent real-time requirements without pretraining, signaling a path for future high-performance, efficient detectors.

Abstract

Enhancing the network architecture of the YOLO framework has been crucial for a long time, but has focused on CNN-based improvements despite the proven superiority of attention mechanisms in modeling capabilities. This is because attention-based models cannot match the speed of CNN-based models. This paper proposes an attention-centric YOLO framework, namely YOLOv12, that matches the speed of previous CNN-based ones while harnessing the performance benefits of attention mechanisms. YOLOv12 surpasses all popular real-time object detectors in accuracy with competitive speed. For example, YOLOv12-N achieves 40.6% mAP with an inference latency of 1.64 ms on a T4 GPU, outperforming advanced YOLOv10-N / YOLOv11-N by 2.1%/1.2% mAP with a comparable speed. This advantage extends to other model scales. YOLOv12 also surpasses end-to-end real-time detectors that improve DETR, such as RT-DETR / RT-DETRv2: YOLOv12-S beats RT-DETR-R18 / RT-DETRv2-R18 while running 42% faster, using only 36% of the computation and 45% of the parameters. More comparisons are shown in Figure 1.

YOLOv12: Attention-Centric Real-Time Object Detectors

TL;DR

This work tackles the tension between speed and accuracy in real-time object detection by infusing YOLO with attention while preserving CNN-style efficiency. It introduces Area Attention (A2) to reduce self-attention cost, Residual Efficient Layer Aggregation Networks (R-ELAN) to stabilize optimization, and targeted architectural refinements (e.g., removing positional encodings, adjusting MLP ratio) to fit the YOLO pipeline. Through extensive COCO experiments across five model scales, YOLOv12 achieves state-of-the-art latency-accuracy trade-offs, outperforming existing real-time detectors and RT-DETR variants on multiple metrics and hardware configurations. The results demonstrate that attention-centric designs can meet stringent real-time requirements without pretraining, signaling a path for future high-performance, efficient detectors.

Abstract

Enhancing the network architecture of the YOLO framework has been crucial for a long time, but has focused on CNN-based improvements despite the proven superiority of attention mechanisms in modeling capabilities. This is because attention-based models cannot match the speed of CNN-based models. This paper proposes an attention-centric YOLO framework, namely YOLOv12, that matches the speed of previous CNN-based ones while harnessing the performance benefits of attention mechanisms. YOLOv12 surpasses all popular real-time object detectors in accuracy with competitive speed. For example, YOLOv12-N achieves 40.6% mAP with an inference latency of 1.64 ms on a T4 GPU, outperforming advanced YOLOv10-N / YOLOv11-N by 2.1%/1.2% mAP with a comparable speed. This advantage extends to other model scales. YOLOv12 also surpasses end-to-end real-time detectors that improve DETR, such as RT-DETR / RT-DETRv2: YOLOv12-S beats RT-DETR-R18 / RT-DETRv2-R18 while running 42% faster, using only 36% of the computation and 45% of the parameters. More comparisons are shown in Figure 1.

Paper Structure

This paper contains 16 sections, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Approach
  4. Efficiency Analysis
  5. Area Attention
  6. Residual Efficient Layer Aggregation Networks
  7. Architectural Improvements
  8. Experiment
  9. Experimental Setup
  10. Comparison with State-of-the-arts
  11. Ablation Studies
  12. Speed Comparison
  13. Diagnosis & Visualization
  14. Conclusion
  15. Limitations
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Comparisons with other popular methods in terms of latency-accuracy (left) and FLOPs-accuracy (right) trade-offs.
  • Figure 2: Comparison of the representative local attention mechanisms with our area attention. Area Attention adopts the most straightforward equal partitioning way to divide the feature map into $l$ areas vertically or horizontally. (default is 4). This avoids complex operations while ensuring a large receptive field, resulting in high efficiency.
  • Figure 3: The architecture comparison with popular modules including (a): CSPNet wang2020cspnet, (b) ELAN wang2022designing_elan, (c) C3K2 (a case of GELAN) wang2024yolov9jocher2024yolov11, and (d) the proposed R-ELAN (residual efficient layer aggregation networks).
  • Figure 4: Comparison with popular methods in terms of accuracy-parameters (left) and accuracy-latency trade-off on CPU (right).
  • Figure 5: Comparison of heat maps between YOLOv10 wang2024yolov10, YOLOv11 jocher2024yolov11, and the proposed YOLOv12. Compared to the advanced YOLOv10 and YOLOv11, YOLOv12 demonstrates a clearer perception of objects in the image. All the results are obtained using the X scale models. Zoom in to compare the details.