This article compiles Meituan's cutting‑edge computer‑vision research and engineering achievements—including CVPR award‑winning segmentation, YOLOv6 releases, GPU inference optimizations, the Food2K dataset, and numerous paper digests—to provide practical insights for visual AI practitioners.
YOLOv6 3.0 raises real‑time object detection performance to a new peak with 57.2% AP and 29 FPS on a T4 GPU, surpassing YOLOv7‑E6E, and introduces RepBi‑PAN Neck, Anchor‑Aided Training, and Decoupled Location Distillation to boost accuracy and efficiency.
The paper proposes a full quantization pipeline for YOLOv6 that combines a re‑parameterization optimizer, partial PTQ, channel‑wise distillation, graph‑scale merging, and GPU‑offloaded preprocessing, enabling an INT8 model to retain ~42 % mAP while delivering over 200 % throughput increase and 40 % QPS gain versus FP16.
The new YOLOv6 2.0 release upgrades lightweight and medium‑large models with a CSPStackRep backbone, self‑distillation, and a custom quantization pipeline, delivering up to 869 FPS for the quantized YOLOv6‑S and achieving 49.5%/52.5% AP on COCO while halving training time.
YOLOv6, developed by Meituan's Vision Intelligence team, introduces a hardware‑friendly backbone, an efficient decoupled head, and advanced training strategies that together achieve up to 35.0% AP at 1242 FPS on COCO while outperforming YOLOv5, YOLOX and other same‑size models across multiple deployment platforms.
