# YOLOX: Exceeding YOLO Series in 2021

**Authors:** Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, Jian Sun  
**arXiv:** [2107.08430](https://arxiv.org/abs/2107.08430)  
**Date:** 2021-07-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2107.08430v2/

## Key points

- Authors present improvements to the YOLO series and introduce a new high-performance detector called YOLOX
- Anchor-free approach and advanced detection techniques used in YOLOX
- Achieves state-of-the-art results across various models
- YOLO-Nano model achieves 25.3% AP on COCO, surpassing NanoDet by 1.8% AP
- Enhanced YOLOv3 achieves 47.3% AP on COCO, outperforming current best practice by 3.0% AP
- Introduce YOLOX-L with similar parameters as YOLOv4-CSP and YOLOv5-L, achieving 50.0% AP on COCO at 68.9 FPS on Tesla V100 GPU, surpassing YOLOv5-L by 1.8% AP
- First place in Streaming Perception Challenge at CVPR 2021 using single YOLOX-L model for autonomous driving applications
- Deploy versions of YOLOX available with support for ONNX, TensorRT NCNN, OpenVINO, and source code on GitHub

## Summary

In this report titled "YOLOX: Exceeding YOLO Series in 2021," authors Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun present their experienced improvements to the YOLO series and introduce a new high-performance detector called YOLOX. The researchers adopt an anchor-free approach for the YOLO detector and employ advanced detection techniques such as a decoupled head and the leading label assignment strategy SimOTA. These enhancements enable them to achieve state-of-the-art results across a wide range of models. The authors evaluate the performance of YOLOX on various models. For instance, they demonstrate that with only 0.91M parameters and 1.08G FLOPs, their YOLO-Nano model achieves an impressive 25.3% AP (average precision) on COCO (Common Objects in Context) dataset, surpassing NanoDet by 1.8% AP. They also enhance the widely used YOLOv3 detector to achieve a remarkable 47.3% AP on COCO, outperforming the current best practice by 3.0% AP. Furthermore, the researchers introduce YOLOX-L which has approximately the same number of parameters as YOLOv4-CSP and YOLOv5-L models. They achieve outstanding results with this model by attaining a 50.0% AP on COCO while maintaining a speed of 68.9 FPS (frames per second) on Tesla V100 GPU; surpassing YOLOv5-L by 1.8% AP. The authors highlight their success in winning the first place in the Streaming Perception Challenge at CVPR 2021 using a single YOLOX-L model for autonomous driving applications. Overall, this report provides valuable insights and experiences for developers and researchers working in practical scenes related to object detection; offering deploy versions of YOLOX with support for ONNX, TensorRT NCNN and OpenVINO as well as source code available on GitHub at https://github/MegviiBaseDetection/Yolox .