# PP-YOLOv2: A Practical Object Detector

**Authors:** Xin Huang, Xinxin Wang, Wenyu Lv, Xiaying Bai, Xiang Long, Kaipeng Deng, Qingqing Dang, Shumin Han, Qiwen Liu, Xiaoguang Hu, Dianhai Yu, Yanjun Ma, Osamu Yoshie  
**arXiv:** [2104.10419](https://arxiv.org/abs/2104.10419)  
**Date:** 2021-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.10419v1/

## Key points

- PP-YOLOv2 is a practical object detector for real-world applications
- The authors conducted a comprehensive evaluation of refinements to improve performance while maintaining inference time
- Incremental ablation study was used to analyze and evaluate the impact of these refinements
- Valuable insights were provided by discussing unsuccessful attempts in the experimentation process
- Multiple effective refinements were combined to significantly increase PP-YOLO's mAP score on COCO2017 test-dev dataset
- PP-YOLOv2 achieves a running speed of 68.9 FPS at an input size of 640x640, which can be further improved with Paddle inference engine, TensorRT, FP16 precision, and batch size 1 resulting in 106.5 FPS
- Compared to other object detectors like YOLOv4-CSP and YOLOv5l with similar parameter counts, PP-YOLOv2 exhibits superior performance in accuracy and efficiency
- When using ResNet101 as the backbone network architecture, it achieves a remarkable mAP score of 50.3% on COCO2017 test-dev dataset
- Source code is available on GitHub for further exploration and implementation

## Summary

PP-YOLOv2 is a practical object detector that aims to be effective and efficient for real-world applications. The authors of the paper conducted a comprehensive evaluation of various refinements to improve the performance of PP-YOLO while keeping the inference time almost unchanged. An incremental ablation study was used to analyze these refinements and empirically evaluate their impact on the model's performance. Additionally, valuable insights were provided into their experimentation process by discussing unsuccessful attempts. By combining multiple effective refinements, they were able to significantly increase PP-YOLO's mAP score from 45.9% to 49.5% on COCO2017 test-dev dataset. In terms of speed, PP-YOLOv2 achieves an impressive running speed of 68.9 frames per second (FPS) at an input size of 640x640 which can be further improved with Paddle inference engine with TensorRT, FP16 precision and batch size 1 resulting in 106.5 FPS. Compared to existing object detectors such as YOLOv4-CSP and YOLOv5l with similar parameter counts, PP-YOLOv2 exhibits superior performance demonstrating its effectiveness and efficiency in object detection tasks. When using ResNet101 as the backbone network architecture, it achieves a remarkable mAP score of 50.3% on COCO2017 test-dev dataset. This paper presents PP-YOLOv2 as a practical object detector that combines multiple effective refinements for state-of-the art performance in terms of accuracy and speed; source code is available on GitHub for further exploration and implementation.