YOLOv6 2.0: Enhanced Object Detection Models and Quantization Solutions

YOLOv6 2.0, the latest version of Meituan’s industrial‑grade single‑stage object detection framework, was released on September 5, 2022. The update introduces a full suite of stronger models, including lightweight YOLOv6‑S (quantized version achieving 869 FPS) and mid‑to‑large models YOLOv6‑M/L with COCO AP of 49.5 % / 52.5 % and inference speeds of 233 FPS / 121 FPS on a T4 GPU (batch size = 32).

Key innovations comprise a new CSPStackRep backbone for the M/L models, systematic evaluation of recent strategies to boost accuracy and speed, a self‑distillation learning scheme, and training‑time reductions of about 50 %.

The release also provides a dedicated quantization pipeline based on RepOptimizer and quantization‑aware training, enabling YOLOv6‑S 2.0 to reach 43.3 mAP at 869 FPS with minimal accuracy loss.

YOLOv6 supports end‑to‑end development (training, evaluation, inference, quantization, distillation) and deployment on diverse platforms such as TensorRT, OpenVINO, ARM (MNN, TNN, NCNN). Detailed tutorials are available in the GitHub repository.

For further technical details, refer to the official Tech Report (arXiv:2209.02976) and the GitHub project at https://github.com/meituan/YOLOv6.