Please reply to the backstage of the public account [ Computer Vision Alliance ] [ 20201118 ] for the pdf of the paper
Summary:
We proved that the YOLOv4 target detection neural network based on the CSP method can be scaled up and down, suitable for small and large networks, while maintaining the best speed and accuracy. We propose a network scaling method, which can not only modify the depth, width, and resolution of the network, but also modify the structure of the network. The YOLOv4 large model achieved the most advanced results: On Tesla V100, the MS COCO data set achieved 55.4% AP (73.3% AP50) at 15 FPS, and as the test time increased, YOLOv4 large reached 55.8% AP (73.2 AP50). As far as we know, this is the highest accuracy currently available on the COCO dataset. The YOLOv4 tiny model achieves 22.0% AP (42.0% AP50) at a speed of 443 FPS on the RTX 2080Ti, and by using TensorRT, batch size=4 and FP16 accuracy, YOLOv4 tiny reaches 1774 FPS.
Feel:
Through the abstract, we think this article is worth reading!
Code related content in Chapter 4 of this article:
YOLOv4-CSP: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp
YOLOv4-tiny: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-tiny
YOLOv4-large: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-large
Updates to this article
Scaled-YOLOv4:
https://github.com/WongKinYiu/ScaledYOLOv4
Overall summary of contributions:
[1] Network expansion improves performance while reducing memory usage
[2] Expanding detection through strategies
[3] Extend the model through relevant content
[4] Designed a model of colleges and universities
Contents of Scaled-YOLOv4:
CSP-ized YOLOv4
YOLOv4-tiny
YOLOv4-large
compare results