Source-linked AI summary
YOLOv6 v3.0: A Full-Scale Reloading
Chuyi Li, Lulu Li, Yifei Geng, Hongliang Jiang, Meng Cheng, Bo Zhang, Zaidan Ke, Xiaoming Xu, Xiangxiang Chu
TL;DR
Real-time object detection requires stronger accuracy without sacrificing inference efficiency. YOLOv6 v3.0 renovates its network and training strategy with neck improvements, anchor-aided training, deeper high-resolution designs, and self-distillation. The upgraded detector achieves state-of-the-art accuracy for real-time object detection, with ablations showing incremental gains from its contributions.
Problem
YOLOv6 seeks higher inference speed as an anchor-free detector, but anchor-based training can provide additional performance gains under the same settings.
Method
The paper combines network redesigns with anchor-aided training, dynamic self-distillation, and an auxiliary DFL regression branch removed after training.
Results
YOLOv6 achieves state-of-the-art accuracy for real-time object detection, while ablations report AP gains of 0.6% from BiC and SimCSPSPPF, 0.3% from AAT, and 0.7% from DLD.
Takeaways & Limitations
The release demonstrates that coordinated network and training upgrades can improve YOLOv6 accuracy while preserving its application-oriented real-time focus.
Takeaways & Limitations
The self-distillation weighting uses the current and maximum training epochs to schedule the balance between losses.
Abstract
from arXiv · showhide
The YOLO community has been in high spirits since our first two releases! By the advent of Chinese New Year 2023, which sees the Year of the Rabbit, we refurnish YOLOv6 with numerous novel enhancements on the network architecture and the training scheme. This release is identified as YOLOv6 v3.0. For a glimpse of performance, our YOLOv6-N hits 37.5% AP on the COCO dataset at a throughput of 1187 FPS tested with an NVIDIA Tesla T4 GPU. YOLOv6-S strikes 45.0% AP at 484 FPS, outperforming other mainstream detectors at the same scale (YOLOv5-S, YOLOv8-S, YOLOX-S and PPYOLOE-S). Whereas, YOLOv6-M/L also achieve better accuracy performance (50.0%/52.8% respectively) than other detectors at a similar inference speed. Additionally, with an extended backbone and neck design, our YOLOv6-L6 achieves the state-of-the-art accuracy in real-time. Extensive experiments are carefully conducted to validate the effectiveness of each improving component. Our code is made available at https://github.com/meituan/YOLOv6.
1. Introduction
YOLOv6 v3.0 renovates the detector’s network design and training strategy to improve real-time object detection. Its enhancements target localization, mixed anchor paradigms, high-resolution detection, and small-model performance.
- YOLOv6 v3.0 renovates both the detector network and training strategy, with comparisons against peers at similar scales.
- The renewed neck adds BiC for more accurate localization signals and SimCSPSPPF for performance gains with negligible speed degradation.
- Anchor-aided training combines anchor-based and anchor-free advantages without changing inference efficiency.
- An additional backbone and neck stage strengthens YOLOv6 for state-of-the-art COCO performance with high-resolution input.
- Self-distillation improves small YOLOv6 models by using a heavier DFL regression branch during training and removing it at inference.
2. Method
YOLOv6 v3.0 renovates the detection neck and training strategy with enhanced feature integration, anchor-aided training, and self-distillation. These additions preserve inference efficiency by using auxiliary branches during training and removing them at inference.
- Detection neck: The enhanced-PAN neck uses a BiC module to fuse feature maps from three adjacent layers, preserving localization signals important for small objects.BiC integrates an extra low-level backbone feature into the corresponding pyramid feature.
- Detection neck: SimCSPSPPF simplifies SPPF into a CSP-like block, while RepBlock or CSPStackRep upgrades and width-depth adjustments tailor the neck to model scale.The stated goal is stronger representation with negligible speed degradation from the SimCSPSPPF change.
- Anchor-aided training: YOLOv6 uses anchor-aided training because anchor-based assignment provides additional performance gains over the anchor-free paradigm on YOLOv6-N under the same settings.Anchor-based ATSS was also used as a warm-up label-assignment strategy in early YOLOv6 versions to stabilize training.
- Anchor-aided training: AAT adds anchor-based auxiliary branches to classification and regression heads, whose independent losses jointly guide the main anchor-free heads during training.The auxiliary branches are removed at inference, so the method is intended to improve accuracy without decreasing speed.
- Self-distillation: Self-distillation applies teacher-student guidance to classification and box regression, combining detection loss with a knowledge-distillation loss.The method uses KL divergence for class predictions and DFL for regression, with the total loss balancing detection and distillation terms.
- Self-distillation: The distillation weight α decays with a cosine schedule across training epochs, shifting emphasis from teacher soft labels toward hard labels as student performance approaches the teacher.Ei denotes the current epoch and Emax the maximum number of training epochs.
3. Experiments
Experiments compare YOLOv6 with YOLO-series detectors and ablate its network and training improvements. The results show gains from BiC, SimCSPSPPF, AAT, and self-distillation across accuracy, localization, and efficiency.
- Comparisons: YOLOv6-L6 improves AP by 0.4% and runs 63% faster than YOLOv7-E6E with batch size 1.
- Ablation Study: The renovated YOLOv6-S network with BiC and SimCSPSPPF improves AP by 0.6%, followed by incremental gains of 0.3% from AAT and 0.7% from DLD.
- Network Design: BiC improves AP by 0.6% on YOLOv6-S and 0.4% on YOLOv6-L when applied to the top-down pathway, with negligible efficiency loss.Applying BiC to the bottom-up pathway provides no positive accuracy gain.
- Network Design: SimCSPSPPF improves AP over SimSPPF by 1.1%/0.4%/0.1% for YOLOv6-N/S/M and runs nearly 10% faster than SimSPPCSPC.SimCSPSPPF is used for YOLOv6-N/S, while SimSPPF is used for YOLOv6-M/L.
- Training Strategy: AAT adds 0.3%/0.5%/0.5% AP for YOLOv6-S/M/L and notably enhances small-object AP for YOLOv6-N/S/M.For YOLOv6-L, AP_l improves further.
- Training Strategy: Self-distillation raises YOLOv6-L by 0.6% AP with weight decay, while doubled training epochs perform worse because of overfitting.
- Training Strategy: DLD gives YOLOv6-S a 0.7% AP boost and performs 0.5% better than training with doubled epochs.
4. Conclusion
YOLOv6 is upgraded through network-design and training-strategy changes. These upgrades boost it to state-of-the-art accuracy for real-time object detection.
- YOLOv6 upgrades its network design and training strategy to achieve state-of-the-art accuracy for real-time object detection.
A.1. Setup
The setup standardizes latency measurement across hardware and software configurations while adding alternate TensorRT and GPU references. This provides a broader view of deployment performance.
- Latency is measured on an NVIDIA Tesla T4 GPU with TensorRT version 7.2.1.6 unless otherwise stated.
- All models are re-measured under the same hardware and software configuration because hardware and software settings have large variance.
- TensorRT version switching and V100 latency measurements provide consistency checks and additional comparison references.
A.2. T4 GPU Latency Table with TensorRT 8
Table 9 compares YOLO-series latency and throughput on a T4 GPU using TensorRT 8.2. YOLOv6 throughput remains comparable to peer models.
- YOLOv6 models maintain throughput comparable to their peers.
- Table 9 compares latency and throughput across YOLO-series models.
- The comparison uses an NVIDIA T4 GPU.
- The measurements use TensorRT version 8.2.
A.3. V100 GPU Latency Table
Table 10 reports that YOLOv6 largely maintains its speed advantage in the V100 GPU comparison.
- YOLOv6 largely maintains its speed advantage in the V100 GPU comparison.
- The reported result concerns YOLO-series speed performance on a V100 GPU.
- Table 10 is the reference for the stated speed comparison.
A.4. CPU Latency
The paper evaluates CPU latency with OpenCV DNN and reports a same-environment YOLO-series latency comparison at FP32 precision and 640×640 input size.
- CPU latency is evaluated on a 2.6 GHz Intel Core i7 using OpenCV Deep Neural Network.
- Table 10 instead reports latency and throughput on a V100 GPU.
- Table 11 compares YOLO-series latency on a typical CPU.
- The CPU comparison uses FP32 precision and 640×640 input sizes in the same environment.