Technical Innovations in YOLOv6 3.0 for Real‑Time Object Detection

1. Overview

On Jan 6 Meituan Visual Intelligence released YOLOv6 3.0. YOLOv6‑L6 achieves 57.2 % AP and 29 FPS on a T4 GPU, surpassing YOLOv7‑E6E and establishing the current real‑time object‑detection SOTA.

Technical report: https://arxiv.org/pdf/2301.05586.pdf. Source code: https://github.com/meituan/YOLOv6.

2. Key Innovations

2.1 RepBi‑PAN Neck

Effective multi‑scale feature fusion is critical for detection. Building on FPN, PANet, BiFPN and PRB‑FPN, a re‑parameterizable bidirectional‑fusion PAN (RepBi‑PAN) is introduced. A Bidirectional Concatenate (BiC) module injects bottom‑up information into the top‑down path, improving shallow‑feature utilization for small‑object localization.

Table 2 (BiC ablation) shows that adding BiC to the top‑down path raises AP by 0.6 % for YOLOv6‑S and 0.4 % for YOLOv6‑L with only a 4 % speed drop, and yields a 1.8 % AP gain for small objects.

Additional experiments compare different SPP modules (Table 3). The simplified SPPCSPC module improves AP by 1.6 % (YOLOv6‑N) and 0.3 % (YOLOv6‑S) but reduces FPS by ~10 %. Replacing SimSPPF with SimCSPSPPF gives 1.1 %/0.4 %/0.1 % AP gains on N/S/M models respectively, with higher inference speed. Consequently SimCSPSPPF is used on YOLOv6‑N/S and SimSPPF on YOLOv6‑M/L.

2.2 Anchor‑Aided Training (AAT)

Anchor‑based and anchor‑free paradigms were evaluated on YOLOv6‑N (Table 4). Overall mAP is similar, but the anchor‑based version attains higher AP on small, medium and large objects.

ATSS pre‑warm stabilizes training: without ATSS, mAP varies 35.9 % → 35.3 % (0.6 % variance); with ATSS, the peak mAP is 35.7 % but variance is reduced.

AAT merges both paradigms: separate anchor‑based and anchor‑free branches compute independent losses that are summed for back‑propagation. A dense sampling mechanism expands candidate boxes and repeats sampling on the same feature point, improving box quality. Auxiliary anchor‑based branches are active only during training, incurring no inference overhead.

Ablation (Table 5) reports a 0.3 % AP gain for YOLOv6‑S and 0.5 % for YOLOv6‑M/L, with notable improvements for small‑object detection.

2.3 Decoupled Location Distillation (DLD)

Traditional Logit Mimicry lacks location‑distillation capability. Adding a DFL branch improves localization but slows small models (YOLOv6‑N speed ↓16.7 %, YOLOv6‑S speed ↓5.2 %). DLD introduces an extra regression branch at each layer that participates in IoU loss and receives guidance from a branch‑distillation scheme using DFL‑generated labels.

During inference the DFL branch is removed, leaving only the strengthened regression branch, thus preserving speed.

Ablation (Table 6) shows that training YOLOv6‑S for 600 epochs yields 44.6 % mAP, while DLD raises it to 45.1 % mAP (+0.5 %) without affecting inference efficiency.

3. Conclusion

YOLOv6 3.0 advances the accuracy‑efficiency trade‑off through RepBi‑PAN, Anchor‑Aided Training and Decoupled Location Distillation, delivering a faster and more precise open‑source framework for industrial object‑detection tasks.