[Paper Review] PP-YOLOE: An evolved version of YOLO

Shangliang Xu, Xinxin Wang+9 more|arXiv (Cornell University)|Mar 30, 2022

Advanced Neural Network Applications224 citations

TL;DR

PP-YOLOE is an industrial one-stage detector that advances PP-YOLOv2 with anchor-free design, CSPRepResNet backbone/neck, ET-head, and TAL label assignment, achieving state-of-the-art COCO results and wide deployment compatibility.

ABSTRACT

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.

Motivation & Objective

• Advance one-stage object detection with a scalable, deployable model family for real-time applications.
• Improve alignment between classification and localization through a task-aligned training paradigm.
• Enhance backbone and neck efficiency with RepResBlock-based CSPRepResNet and CSPRepResStage.
• Incorporate dynamic label assignment and a fast, efficient head to boost accuracy without heavy deformable ops or Matrix NMS.

Proposed method

• Adopt anchor-free detection inspired by FCOS with pixel-level positive sampling.
• Introduce RepResBlock for backbone/neck and CSPRepResNet/CSPRepResStage for efficient cross-stage connections.
• Incorporate Task Alignment Learning (TAL) with dynamic label assignment to improve positive sample selection and localization.
• Develop Efficient Task-aligned Head (ET-head) combining ESE attention, VFL/DFL-based losses, and distribution-based bounding box modeling.
• Scale models with width/depth multipliers to produce s/m/l/x variants and enable flexible deployment.
• Provide deployment-ready implementation with PaddleDetection, TensorRT and ONNX support.

Experimental results

Research questions

• RQ1 Does anchor-free training with TAL provide measurable gains over anchor-based PP-YOLOv2 and competitive models like YOLOX?
• RQ2 What is the impact of the CSPRepResNet/RepResBlock backbone-neck design on speed and accuracy across model scales?
• RQ3 How does ET-head with VFL/DFL losses affect classification-localization alignment and overall AP?
• RQ4 Can the proposed PP-YOLOE series meet industrial deployment constraints (speed, memory, hardware compatibility) across variants?

Key findings

• PP-YOLOE-l achieves 51.4 mAP on COCO test-dev with 78.1 FPS on Tesla V100.
• PP-YOLOE-l improves AP by 1.9 or 1.3 points compared to PP-YOLOv2 and YOLOX respectively, while also boosting speed.
• The anchor-free design reduces reliance on hand-tuned anchors with minimal AP loss (0.3) relative to baseline PP-YOLOv2.
• TAL provides its best 45.2 AP in ablations, and integrating TAL yields 50.4 AP as baseline when combined with other components.
• ET-head adds 0.5 AP over the base and maintains favorable speed, contributing to 50.9 AP.
• PP-YOLOE-superior models (x) reach 52.2–54.7 AP with 95.2–149.2 FPS depending on precision and deployment hardware.
• PP-YOLOE-family achieves competitive or superior speed/accuracy trade-offs vs YOLOv5/YOLOX across FP32/FP16/TensorRT configurations.