YOLOv6 2.0: Enhanced Object Detection Models and Quantization Solutions

Release Overview

YOLOv6 2.0, an open‑source single‑stage object detection framework for industrial applications, adds a full family of models: lightweight quantized YOLOv6‑S and medium/large YOLOv6‑M/L.

Performance Highlights

On COCO validation, YOLOv6‑M achieves 49.5 % AP, YOLOv6‑L 52.5 % AP. Inference on an NVIDIA T4 GPU with TensorRT FP16, batch = 32, yields 233 FPS (M) and 121 FPS (L). The quantized YOLOv6‑S reaches 43.3 % mAP and 869 FPS under the same conditions.

Model Improvements

CSPStackRep backbone replaces the previous Single Path design, delivering higher overall performance for M/L models.

Systematic evaluation of recent strategies across model sizes selected the best accuracy‑speed trade‑off and reduced total training time by 50 %.

Self‑distillation with a novel learning schedule substantially boosts M/L accuracy.

Early‑stop strong data augmentation during training and image‑resize optimization at inference fix the precision loss caused by forced 640×640 inputs.

Quantization Scheme

The quantization pipeline builds on RepOptimizer (arXiv:2205.15242) to re‑parameterize gradients, mitigating the large dynamic range of multi‑branch layers. It supports post‑training quantization (PTQ) with minimal accuracy drop and a channel‑wise distillation‑aware quantization‑aware training (QAT) that further improves precision. Compared with PaddleSlim, the YOLOv6‑S 2.0 quantization delivers higher mAP at similar speed.

Deployment Support

YOLOv6 provides end‑to‑end tooling for training, evaluation, inference, quantization, and distillation. Deployment targets include GPU (TensorRT), CPU (OpenVINO), and ARM runtimes (MNN, TNN, NCNN). Detailed instructions reside in the GitHub repository’s Deployment directory.

References

GitHub repository: https://github.com/meituan/YOLOv6

Technical report (arXiv): https://arxiv.org/abs/2209.02976