Yolov3, 4, 5 and 6 Summary of target detection

Background description

cv A typical problem in is target detection , In target detection yolo The model is a classical object detection model , In various target detection competitions and data sets , The performance is very bright , Borrow this article to yolov3、v4、v5、tph_v5、v6 Make a summary .

yolo-v3、v4

The alchemy furnace ：darknet
darknet git Address ：https://github.com/AlexeyAB/darknet
Usage plan ： Replaceable cfg Represents the network structure switching v3 And v4 To train different models , The adjustment of super parameters is also cfg In the document .

yolov5

The alchemy furnace ：pytorch
git Address ：https://github.com/ultralytics/yolov5
Usage plan ： have access to wandb and tensorboard To train visualization , Than v3、v4 Better compare the accuracy of each version of the model

tph_yolov5

The alchemy furnace ：pytorch
git Address ：https://github.com/cv516Buaa/tph-yolov5
advantage ： Optimized for large images and small targets , By looking at the pictures during the training , It is necessary to eliminate the confidence , Sensitive to small targets , Others are basically v5 Updated version of , Take out the super parameters as a separate file , When training, you can use .

yolov6

The alchemy furnace ：pytorch
git Address ：https://github.com/meituan/YOLOv6
advantage ： The experimental data looks beautiful , The optimized scheme is the accumulation of various optimization strategies , There is no obvious optimization , But the data is good , Made by meituan , There is no paper support .