1. Overview  

YOLOv12 introduces an **attention-centric architecture**, departing from the traditional CNN-based approaches used in previous YOLO models, while still maintaining the **real-time inference speed** required by many practical applications. Through novel innovations in attention mechanisms and overall network design, YOLOv12 achieves **state-of-the-art object detection accuracy** without compromising real-time performance.

2. Key Features  

**Regional Attention Mechanism**:  

A new self-attention method designed for efficiently handling large receptive fields. It divides the feature map into *l* equal-sized regions (default: 4) either horizontally or vertically, avoiding computationally expensive operations while preserving a large effective receptive field. This significantly reduces computational cost compared to standard self-attention.

**Residual Efficient Layer Aggregation Network (R-ELAN)**:  

An enhanced feature aggregation module based on ELAN, specifically designed to address optimization challenges in large-scale, attention-centric models. Key improvements include:  

- Block-level residual connections with scaling (similar to LayerScale).  

- A redesigned feature aggregation strategy that creates bottleneck-like structures for better efficiency.

**Optimized Attention Architecture**:  

YOLOv12 streamlines the standard attention mechanism for higher efficiency and seamless integration into the YOLO framework:  

- Employs **FlashAttention** to minimize memory access overhead.  

- **Removes positional encoding**, resulting in a simpler and faster model.  

- Adjusts the MLP expansion ratio (from the typical 4× down to **1.2× or 2×**) to better balance computation between attention and feed-forward layers.  

- Reduces the depth of stacked blocks to improve trainability.  

- Strategically incorporates **convolutional operations** to boost computational efficiency.  

- Adds a **7×7 depthwise separable convolution** (“position-aware module”) within the attention block to implicitly encode spatial information.

**Comprehensive Task Support**:  

YOLOv12 supports a wide range of core computer vision tasks:  

- Object Detection  

- Instance Segmentation  

- Image Classification  

- Pose Estimation  

- Oriented Bounding Box (OBB) Detection  

**Higher Efficiency**:  

YOLOv12 achieves **higher accuracy with fewer parameters** than many prior models, striking an exceptional balance between speed and precision.

**Flexible Deployment**:  

Designed for deployment across diverse platforms—from **edge devices** to **cloud infrastructure**—ensuring high performance in resource-constrained environments.

*(Visualization: YOLOv12 comparison chart)*

3. Supported Tasks and Modes  

YOLOv12 supports multiple computer vision tasks. The table below outlines task coverage and supported operational modes (Inference, Validation, Training, Export):

| Model Type     | Task               | Inference | Validation | Training | Export |

|----------------|--------------------|-----------|------------|----------|--------|

| YOLOv12        | Detection          | ✅        | ✅         | ✅       | ✅     |

| YOLOv12-seg    | Segmentation       | ✅        | ✅         | ✅       | ✅     |

| YOLOv12-pose   | Pose Estimation    | ✅        | ✅         | ✅       | ✅     |

| YOLOv12-cls    | Classification     | ✅        | ✅         | ✅       | ✅     |

| YOLOv12-obb    | OBB Detection      | ✅        | ✅         | ✅       | ✅     |

 4. Performance Evaluation  

Evaluated on the **COCO val2017** dataset, YOLOv12 demonstrates outstanding performance across all model scales (input size: 640×640):

| Model       | mAP (%) | Latency (ms) | Parameters | FLOPs (G) |

|-------------|---------|--------------|------------|-----------|

| YOLOv12-N   | 40.6    | 1.64         | 2.6M       | 6.5       |

| YOLOv12-S   | 48.0    | 2.61         | 9.3M       | 21.4      |

| YOLOv12-M   | 52.5    | 4.86         | 20.2M      | 67.5      |

| YOLOv12-L   | 53.7    | 6.77         | 26.4M      | 88.9      |

| YOLOv12-X   | 55.2    | 11.79        | 59.1M      | 199.0     |

Compared to earlier versions (e.g., YOLOv10 and YOLOv11), YOLOv12 shows **significant accuracy gains** with comparable speed. For example:  

- YOLOv12-N improves mAP by **+2.1%** over YOLOv10-N and **+1.2%** over YOLOv11-N, with similar latency.  

- Similar advantages are consistently observed across other model sizes.

5. Comprehensive Multi-Task Support  

Beyond object detection, YOLOv12 excels in **instance segmentation, image classification, pose estimation, and oriented object detection (OBB)**. This versatility makes it highly adaptable across diverse real-world applications.

6. Flexible Deployment Capability  

Engineered for **cross-platform deployment**, YOLOv12 runs efficiently on everything from **low-power edge devices** to **high-performance cloud servers**. Its optimized compute and memory footprint enable high-accuracy inference even under strict hardware constraints.

7. Conclusion  

YOLOv12 represents a major leap forward in real-time object detection. By integrating an **attention-centric architecture**, **R-ELAN**, and a suite of **optimized attention techniques**, it achieves **simultaneous improvements in both accuracy and speed**.  

Compared to previous YOLO generations, YOLOv12 delivers **measurable gains across all metrics**, particularly in maintaining real-time inference while significantly boosting detection performance. Coupled with its **multi-task support** and **deployment flexibility**, YOLOv12 is poised to become a powerful tool for both research and industrial applications.

In summary, the release of YOLOv12 marks another milestone in real-time vision AI—offering a more capable, efficient, and versatile foundation for the next generation of intelligent systems.