# YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

**Authors:** Chien-Yao Wang, Alexey Bochkovskiy, Hong-Yuan Mark Liao  
**arXiv:** [2207.02696](https://arxiv.org/abs/2207.02696)  
**Date:** 2022-07-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.02696v1/

## Key points

- YOLOv7 is a state-of-the-art real-time object detector
- It surpasses all known detectors in terms of speed and accuracy
- Achieves highest accuracy of 56.8% AP among real-time object detectors with frame rate of 30 FPS or higher on GPU V100
- YOLOv7-E6 variant performs even better with frame rate of 56 FPS on V100 and accuracy of 55.9% AP
- Outperforms transformer-based SWIN-L Cascade-Mask R-CNN by 509% in speed and 2% in accuracy
- Outperforms convolutional-based ConvNeXt-XL Cascade-Mask R-CNN by 551% in speed and 0.7% AP in accuracy
- Outperforms other well-known object detectors such as YOLOR, YOLOX, Scaled-YOLOv4, YOLOv5, DETR, Deformable DETR, DINO-5scale-R50, ViT Adapter B, etc.
- Trained solely on MS COCO dataset from scratch without using any other datasets or pre-trained weights
- Source code freely available on GitHub at https://github.com/WongKinYiu/yolov7

## Summary

YOLOv7 is a state-of-the-art real-time object detector that surpasses all known detectors in terms of both speed and accuracy. It achieves this by achieving the highest accuracy of 56.8% AP among all real-time object detectors with a frame rate of 30 FPS or higher on the GPU V100. The YOLOv7-E6 variant of the detector performs even better, with a frame rate of 56 FPS on the V100 and an accuracy of 55.9% AP. Comparing YOLOv7-E6 to other popular detectors, it outperforms the transformer-based SWIN-L Cascade-Mask R-CNN by 509% in speed and 2% in accuracy (with a frame rate of 9.2 FPS on the A100 and an accuracy of 53.9% AP). Similarly, it outperforms the convolutional-based ConvNeXt-XL Cascade-Mask R-CNN by 551% in speed and 0.7% AP in accuracy (with a frame rate of 8.6 FPS on the A100 and an accuracy of 55.2% AP). In addition to these specific detectors, YOLOv7 also outperforms other well-known object detectors such as YOLOR, YOLOX, Scaled-YOLOv4, YOLOv5, DETR, Deformable DETR, DINO-5scale-R50, ViT Adapter B and many more in terms of both speed and accuracy. What sets YOLOv7 apart is that it is trained solely on the MS COCO dataset from scratch without using any other datasets or pre-trained weights; demonstrating its ability to achieve impressive performance without relying on external resources. The source code for YOLOv7 is freely available on GitHub at https://github.com/WongKinYiu/yolov7 allowing researchers and developers to access and utilize this advanced object detection system.