# YOLOv3: An Incremental Improvement

**Authors:** Joseph Redmon, Ali Farhadi  
**arXiv:** [1804.02767](https://arxiv.org/abs/1804.02767)  
**Date:** 2018-04-08  
**Summary page:** https://www.summarizepaper.com/en/paper/1804.02767v1/

## Key points

- YOLOv3 is an updated version of the YOLO algorithm
- The authors made design changes to improve performance and trained a new network
- The new YOLOv3 network is more accurate and slightly larger than the previous version, but still fast
- At a resolution of 320x320 pixels, it runs in just 22 milliseconds with an mAP of 28.2
- YOLOv3 performs well when considering the old .5 IOU mAP detection metric
- YOLOv3 achieves an mAP@50 of 57.9 in 51 milliseconds on a Titan X GPU
- Compared to RetinaNet, YOLOv3 offers similar performance but is significantly faster (3.8 times)
- Code for YOLOv3 is available online at https://pjreddie.com/yolo/
- YOLOv3 combines impressive performance metrics with efficient runtime speeds

## Summary

The paper titled "YOLOv3: An Incremental Improvement" by Joseph Redmon and Ali Farhadi presents updates to the YOLO (You Only Look Once) algorithm. The authors made several design changes to improve its performance and trained a new network that is more accurate and slightly larger than the previous version. Despite being larger, the new YOLOv3 network remains fast. At a resolution of 320x320 pixels, it runs in just 22 milliseconds with an average precision (mAP) of 28.2, which is comparable to SSD but three times faster. The authors highlight that YOLOv3 performs well when considering the old .5 IOU mAP detection metric. In terms of specific metrics, YOLOv3 achieves an mAP@50 of 57.9 in 51 milliseconds on a Titan X GPU. In comparison, RetinaNet achieves an mAP@50 of 57.5 in 198 milliseconds on the same hardware; this demonstrates that YOLOv3 offers similar performance to RetinaNet but is significantly faster with a speed improvement of 3.8 times. The paper concludes by mentioning that all the code for YOLOv3 is available online at https://pjreddie.com/yolo/, providing users with easy access to this improved algorithm and its associated benefits such as enhanced accuracy while maintaining its speed and efficiency. Overall, this incremental improvement to the YOLO algorithm makes it a powerful tool for object detection tasks due to its impressive performance metrics combined with its efficient runtime speeds.