Learning-Deep-Learning

YOLOv5 by Ultralytics

December 2020

tl;dr: Pytorch-native improved version of Yolov4.

Overall impression

The author of the repo has not released a paper yet, but the repo is highly useful. Many Kaggler commented that its performance is better than yolov4. The training speed of Yolov5 is also much faster than Yolov4.

Key ideas

• Two great reviews on Zhihu
  • 使用YOLO V5训练自动驾驶目标检测网络
  • 深入浅出Yolo系列之Yolov5核心基础知识完整讲解
• Focus layer
  • This is a spatial to channel layer that warps H x W x 3 into H/2 x W/2 x12.
  • See issue in yolov5 github
  • Papers such as TResNet WACV 2021 and Isometric Neural Networks ICCV 2019 workshop
• Adaptive anchor learning with genetic algorithm

Technical details

• Mosaic data aug was first invented in ultralytics’s yolov3 and borrowed into Yolov4.

Notes

• Questions and notes on how to improve/revise the current work