# YOLOv4: Optimal Speed and Accuracy of Object Detection

**Authors:** Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao  
**arXiv:** [2004.10934](https://arxiv.org/abs/2004.10934)  
**Date:** 2020-04-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.10934v1/

## Key points

- The paper discusses features that improve the accuracy of CNNs for object detection tasks
- Practical testing on large datasets and theoretical justification are emphasized
- Universal features include WRC, CSP, CmBN, SAT, and Mish-activation
- New features introduced are Mosaic data augmentation, DropBlock regularization, and CIoU loss
- Combination of these features achieves state-of-the-art results in object detection
- Results include an AP of 43.5% with an AP50 score of 65.7% on the MS COCO dataset at a real-time speed of approximately 65 frames per second on Tesla V100
- Source code available on GitHub at https://github.com/AlexeyAB/darknet

## Summary

The paper titled "YOLOv4: Optimal Speed and Accuracy of Object Detection" by Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao discusses the various features that can improve the accuracy of Convolutional Neural Networks (CNNs) for object detection tasks. The authors emphasize the need for practical testing of these features on large datasets and theoretical justification of their results. While some features are specific to certain models or problems, others like batch-normalization and residual-connections are applicable to a wide range of models, tasks, and datasets. The authors propose that universal features include Weighted-Residual-Connections (WRC), Cross-Stage-Partial-connections (CSP), Cross mini-Batch Normalization (CmBN), Self-adversarial-training (SAT), and Mish-activation. In addition to these universal features, the authors introduce new features such as Mosaic data augmentation, DropBlock regularization, and CIoU loss. They combine these features with the universal ones to achieve state-of-the-art results in object detection. Specifically, they report an average precision (AP) of 43.5% with an AP50 score of 65.7% for the MS COCO dataset at a real-time speed of approximately 65 frames per second on Tesla V100. The source code for implementing these techniques is available on GitHub at https://github.com/AlexeyAB/darknet. Overall, this paper provides valuable insights into improving CNN accuracy for object detection tasks through a combination of existing and novel features.