sunsmarterjie/yolov12

What it solves

YOLOv12 addresses the speed-accuracy trade-off in real-time object detection. While attention mechanisms typically offer superior modeling capabilities compared to CNNs, they are usually too slow for real-time use. YOLOv12 provides an attention-centric framework that matches the speed of CNN-based detectors while harnessing the performance benefits of attention.

How it works

It implements an attention-centric architecture designed for high-efficiency real-time performance. The project provides multiple model scales (Nano, Small, Medium, Large, and X-Large) and supports three primary computer vision tasks: object detection, instance segmentation, and image classification. It is built upon the Ultralytics framework, allowing for easy training, validation, and export to formats like TensorRT and ONNX.

Who it’s for

Computer vision engineers and researchers who need high-accuracy, real-time object detection, segmentation, or classification on GPUs (such as the T4) or edge devices.

Highlights

• Attention-Centric Design: Combines the accuracy of attention mechanisms with the speed of CNNs.
• Multi-Task Support: Includes dedicated models for detection, instance segmentation, and classification.
• High Efficiency: Outperforms previous YOLO versions (v10, v11) and RT-DETR in accuracy and speed trade-offs.
• Turbo Version: Offers a "Turbo" variant for even faster inference.
• Flexible Deployment: Supports export to TensorRT and ONNX for deployment on Android or other edge devices.