# PP-YOLO: An Effective and Efficient Implementation of Object Detector

**Authors:** Xiang Long, Kaipeng Deng, Guanzhong Wang, Yang Zhang, Qingqing Dang, Yuan Gao, Hui Shen, Jianguo Ren, Shumin Han, Errui Ding, Shilei Wen  
**arXiv:** [2007.12099](https://arxiv.org/abs/2007.12099)  
**Date:** 2020-07-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.12099v3/

## Key points

- PP-YOLO is an effective and efficient implementation of an object detector in computer vision applications.
- The balance between effectiveness and efficiency is crucial, especially considering hardware limitations.
- PP-YOLO builds upon the YOLOv3 model, combining existing tricks to improve accuracy without significantly increasing parameters and FLOPs.
- Experiments are conducted using PaddlePaddle, leading to the name PP-YOLO.
- PP-YOLO achieves a better balance between effectiveness (measured by mAP) and efficiency (measured by FPS).
- PP-YOLO outperforms state-of-the-art detectors like EfficientDet and YOLOv4 with an mAP of 45.2% and an FPS of 72.9.
- This refined implementation strikes a good balance between accuracy and speed, making it highly applicable in practical scenarios.
- Source code for PP-YOLO can be found at https://github.com/PaddlePaddle/PaddleDetection.

## Summary

PP-YOLO is an effective and efficient implementation of an object detector that plays a crucial role in computer vision applications. The balance between effectiveness and efficiency is essential for object detection, especially considering hardware limitations. This paper proposes to build upon the widely used YOLOv3 model to create PP-YOLO which focuses on combining existing tricks that do not significantly increase the number of model parameters and FLOPs (floating-point operations) to improve the accuracy of the detector while maintaining almost unchanged speed. All experiments are conducted using PaddlePaddle, leading to the name PP-YOLO. By leveraging multiple techniques, PP-YOLO achieves a better balance between effectiveness (measured by mean Average Precision or mAP) and efficiency (measured by Frames Per Second or FPS). The results show that PP-YOLO outperforms state-of-the-art detectors like EfficientDet and YOLOv4 with an mAP of 45.2% and an FPS of 72.9. Thus, this paper presents a refined implementation of an object detector that strikes a good balance between accuracy and speed, making it highly applicable in practical scenarios. The source code for PP-YOLO can be found at https://github.com/PaddlePaddle/PaddleDetection.