Abstract

In this report, we present PP-YOLOE, an industrial state-of-the-art objectdetector with high performance and friendly deployment. We optimize on thebasis of the previous PP-YOLOv2, using anchor-free paradigm, more powerfulbackbone and neck equipped with CSPRepResStage, ET-head and dynamic labelassignment algorithm TAL. We provide s/m/l/x models for different practicescenarios. As a result, PP-YOLOE-l achieves 51.4 mAP on COCO test-dev and 78.1FPS on Tesla V100, yielding a remarkable improvement of (+1.9 AP, +13.35% speedup) and (+1.3 AP, +24.96% speed up), compared to the previous state-of-the-artindustrial models PP-YOLOv2 and YOLOX respectively. Further, PP-YOLOE inferencespeed achieves 149.2 FPS with TensorRT and FP16-precision. We also conductextensive experiments to verify the effectiveness of our designs. Source codeand pre-trained models are available athttps://github.com/PaddlePaddle/PaddleDetection.