YOLOv12 — Attention-Centric Real-Time Object Detection

Paper: YOLOv12: Attention-Centric Real-Time Object Detectors (NeurIPS 2025)
Official Repo: github.com/sunsmarterjie/yolov12

---

What is YOLOv12?

YOLOv12 is the first attention-based YOLO model that matches the speed of CNN-based detectors while leveraging the accuracy benefits of attention mechanisms. It introduces two key innovations:

• Area Attention (A²) — splits feature maps into local segments (horizontal/vertical) to reduce the quadratic complexity of full self-attention, keeping inference fast.
• R-ELAN (Residual Efficient Layer Aggregation Network) — adds scaled residual connections for better training stability, especially on larger model scales.

Performance Highlights

Model	mAP (COCO)	Latency (T4 GPU)
YOLOv12-N	40.6%	1.64 ms
YOLOv12-S	48.0%	2.61 ms
YOLOv12-M	52.5%	4.86 ms
YOLOv12-L	53.7%	6.77 ms
YOLOv12-X	55.2%	11.79 ms

YOLOv12-N outperforms YOLOv11-N by +1.2% mAP at comparable speed.

---

Requirements

• Python 3.11
• CUDA-compatible GPU (Turing, Ampere, Ada Lovelace, or Hopper recommended for FlashAttention)
  • Examples: RTX 20/30/40 series, A100, H100, T4
• PyTorch 2.2+

⚠️ Older GPU architectures (e.g., GTX 10 series) may not support FlashAttention. Training will still work but at reduced speed.

---

Installation

# Clone the repo
git clone https://github.com/sunsmarterjie/yolov12.git
cd yolov12

# Create a conda environment
conda create -n yolov12 python=3.11
conda activate yolov12

# Install FlashAttention (for supported GPUs)
wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu11torch2.2cxx11abiFALSE-cp311-cp311-linux_x86_64.whl

# Install dependencies
pip install -r requirements.txt
pip install -e .

---

Dataset Structure

dataset/
├── images/
│   ├── train/
│   └── val/
└── labels/
    ├── train/
    └── val/

Each label .txt file (YOLO format):

<class_id> <x_center> <y_center> <width> <height>

All values normalized between 0.0 and 1.0.

dataset.yaml

path: ./dataset
train: images/train
val: images/val

nc: 3
names:
  0: cat
  1: dog
  2: bird

---

Training

# Train from CLI
yolo train model=yolov12n.pt data=dataset.yaml epochs=100 imgsz=640

# Or with Python
from ultralytics import YOLO

model = YOLO('yolov12n.pt')
model.train(data='dataset.yaml', epochs=100, imgsz=640, batch=16, device=0)

Model Variants

Model	Use Case
yolov12n	Fastest, laptop/edge GPU
yolov12s	Balanced speed & accuracy
yolov12m	Medium accuracy
yolov12l	High accuracy
yolov12x	Best accuracy, needs VRAM

Laptop GPU Tips

VRAM	Recommended Settings
4 GB	batch=4, imgsz=416, yolov12n
6 GB	batch=8, imgsz=512, yolov12s
8 GB+	batch=16, imgsz=640, yolov12m

---

Validation

yolo val model=runs/train/exp/weights/best.pt data=dataset.yaml

model = YOLO('runs/train/exp/weights/best.pt')
metrics = model.val(data='dataset.yaml')
print("mAP50:", metrics.box.map50)
print("mAP50-95:", metrics.box.map)

---

Inference

# Image
yolo predict model=yolov12n.pt source=image.jpg conf=0.5

# Webcam
yolo predict model=yolov12n.pt source=0 show=True

model = YOLO('yolov12n.pt')
results = model.predict('image.jpg', conf=0.5, save=True)

---

Export

yolo export model=yolov12n.pt format=onnx

Supported formats: onnx, torchscript, tflite, engine (TensorRT)

---

Project Structure

yolov12-training/
├── dataset/
│   ├── images/
│   └── labels/
├── configs/
│   └── dataset.yaml
├── scripts/
│   ├── train.py
├── runs/              # Training outputs (auto-generated)
├── requirements.txt
└── README.md

---

License

This project follows the license of the official YOLOv12 repository.