PaperlayerView exact revision on arXiv

Source-linked AI summary

PP-YOLOE: An evolved version of YOLO

Shangliang Xu, Xinxin Wang, Wenyu Lv, Qinyao Chang, Cheng Cui, Kaipeng Deng, Guanzhong Wang, Qingqing Dang, Shengyu Wei, Yuning Du, Baohua Lai

arXiv:2203.16250v3cs.CV

TL;DR

PP-YOLOE addresses the need for high-performing one-stage detection with deployment-friendly speed and accuracy. It evolves PP-YOLOv2 using an anchor-free design, stronger backbone and neck, ET-head, and TAL, while offering scalable models. PP-YOLOE-l reaches 51.4 mAP and 78.1 FPS on the reported COCO and Tesla V100 evaluation.

  • Problem

    One-stage detectors seek a strong speed–accuracy trade-off for real-time applications, while existing designs motivate further optimization of PP-YOLOv2.

    Source: §1 · p. 1

  • Method

    PP-YOLOE combines an anchor-free paradigm, CSPRepResNet-based backbone and neck, ET-head, dynamic label assignment with TAL, and scalable model configurations.

    Source: §2.2 · pp. 2–4, §4 · p. 5

  • Results

    51.4 mAP on COCO test-dev and 78.1 FPS on Tesla V100 are achieved by PP-YOLOE-l, surpassing PP-YOLOv2 by 1.9 AP and YOLOX-l by 1.3% AP.

    Source: §1 · p. 1

  • Takeaways & Limitations

    PP-YOLOE provides s/m/l/x detector variants aimed at different practical scenarios and supports deployment through PaddlePaddle.

    Source: §1 · p. 1, §4 · p. 5

Abstract

from arXiv · show

In this report, we present PP-YOLOE, an industrial state-of-the-art object detector with high performance and friendly deployment. We optimize on the basis of the previous PP-YOLOv2, using anchor-free paradigm, more powerful backbone and neck equipped with CSPRepResStage, ET-head and dynamic label assignment algorithm TAL. We provide s/m/l/x models for different practice scenarios. As a result, PP-YOLOE-l achieves 51.4 mAP on COCO test-dev and 78.1 FPS on Tesla V100, yielding a remarkable improvement of (+1.9 AP, +13.35% speed up) and (+1.3 AP, +24.96% speed up), compared to the previous state-of-the-art industrial models PP-YOLOv2 and YOLOX respectively. Further, PP-YOLOE inference speed achieves 149.2 FPS with TensorRT and FP16-precision. We also conduct extensive experiments to verify the effectiveness of our designs. Source code and pre-trained models are available at https://github.com/PaddlePaddle/PaddleDetection.

1. Introduction

PP-YOLOE evolves PP-YOLOv2 into a scalable detector targeting a stronger speed–accuracy trade-off and broader hardware deployment. Its l model improves reported COCO performance and V100 speed over prior industrial baselines.

Reported results

  • 51.4 mAP and 78.1 FPS are achieved by PP-YOLOE-l on COCO test-dev using a Tesla V100.Source: Figure 1 · p. 1, §1 · p. 1
  • 1.9 AP and 9.2 FPS improvements are reported over PP-YOLOv2 for PP-YOLOE-l.Source: Figure 1 · p. 1

Motivation and contribution

  • PP-YOLOE is proposed as an evolved version of PP-YOLOv2 for improved speed–accuracy trade-offs in one-stage detection.The work builds on prior YOLO-family detectors, including YOLOX and PP-YOLOv2.Source: §1 · p. 1

Deployment

  • PP-YOLOE avoids deformable convolution and Matrix NMS to support deployment across various hardware.Source: §1 · p. 1
  • The model family scales across practical scenarios through s/m/l/x variants and configurable architecture settings.The series is intended for hardware with different computing power.Source: §1 · p. 1
  • PaddleDetection provides the released implementation with TensorRT and ONNX support.Source: §1 · p. 2

2. Method

PP-YOLOE evolves PP-YOLOv2 through an anchor-free design, scalable backbone-neck architecture, task-aligned head, and dynamic label assignment. Its components target accuracy, efficiency, and deployment across varied hardware and scenarios.

Label assignment

  • PP-YOLOE replaces PP-YOLOv2’s anchor-based assignment with an anchor-free method that selects positive samples using feature-map bounds and box-center proximity.The anchor mechanism is described as introducing hyperparameters and relying on hand-crafted design.Source: §2.2 · p. 2

Scalability

  • Width multiplier α and depth multiplier β jointly scale the backbone and neck, producing networks with different parameter counts and computation costs.The paper uses these multipliers to obtain a series of detection networks for different model specifications.Source: §2.2 · p. 2, Table 1 · p. 2

Network structure

  • PP-YOLOE combines CSPRepResNet, PAN, and ET-head, with RepResBlock and CSPRepResStage used to construct the backbone and neck.RepResBlock combines residual and dense connections and can be re-parameterized into a basic residual block for inference.Source: §2.2 · p. 2, Figure 2 · p. 3

Label assignment

  • TAL uses prediction-aware dynamic label assignment and task-aligned loss to address classification-localization misalignment.Dynamic assignment allocates a variable number of positive anchors for each ground-truth object.Source: §2.2 · p. 3

3. Experiment

The experiments train PP-YOLOE on MS COCO-2017 and evaluate ablations on validation data and final results on test-dev. Detector speeds are compared at batch size 1 under specified precision and deployment settings, with additional TensorRT FP16 testing for fairness.

  • 118k MS COCO-2017 training images across 80 classes support the experiments, while ablations use the 5000-image validation set and final results use test-dev.Source: §3 · p. 4
  • 300 training epochs use SGD with momentum 0.9, weight decay 5e-4, cosine learning-rate scheduling, five warmup epochs, and a base learning rate of 0.01.Training uses a total batch size of 64 across 8 × 32 G V100 GPUs and EMA decay 0.9998.Source: §3.1 · p. 4
  • Inference speed is measured with batch size 1 without data preprocessing or non-maximum suppression, while PP-YOLOE uses the Paddle inference engine.Source: §3.2 · p. 4
  • FP16 TensorRT 6.0 speed is additionally tested in the same environment to provide a fair deployment comparison.PaddlePaddle2 officially supports TensorRT for model deployment.Source: §3.2 · p. 4
  • Table 4 compares different detectors on COCO 2017 test-dev, distinguishing updated official results, independently tested results, and precision settings.The table uses FP32 by default without TensorRT and FP16 with TensorRT; YOLOv5 is omitted because its tested input size is not exactly 640 × 640.Source: Table 4 · p. 5

4. Conclusion

PP-YOLOE updates PP-YOLOv2 with scalable architecture, task-aligned prediction, label assignment, and loss designs, producing models intended for varied practical scenarios and deployment.

  • PP-YOLOE combines a scalable backbone-neck architecture, efficient task-aligned head, advanced label assignment, and refined objective loss.Source: §4 · p. 5
  • The s/m/l/x model series is designed to cover different practical scenarios and supports smooth deployment through PaddlePaddle.Source: §4 · p. 5

Related papers

Loading 2203.16250v3…