YOLOv3是2018年出的,当时保存了一些内容,点这里
现在继续更新其他人写的:
1 重磅更新!YoloV4最新论文与源码!权重!结构!翻译!
3 未完待续
它的配置可以参考yolov3,在win10+vs2015+CUDA 9.0,是可以运行的。CUDNN+opencv3.2没有问题
首先用yolov4.cfg 和yolov4.weights来简单测试图片,效果如下
person: 95%
person: 63%
person: 98%
handbag: 53%
person: 78%
backpack: 61%
car: 43%
traffic light: 91%
person: 76%
car: 44%
person: 50%
person: 55%
handbag: 54%
person: 99%
handbag: 26%
handbag: 28%
handbag: 32%
person: 98%
handbag: 48%
person: 98%
person: 61%
person: 70%
backpack: 45%
person: 83%
person: 75%
person: 69%
person: 72%
handbag: 33%
handbag: 55%
person: 98%
car: 64%
car: 98%
person: 79%
person: 68%
handbag: 39%
person: 37%
person: 34%
car: 98%
person: 58%
person: 34%
car: 31%
person: 67%
person: 40%
person: 55%
person: 93%
backpack: 61%
handbag: 26%
car: 92%
person: 83%
car: 41%
person: 30%
car: 99%
person: 71%
car: 78%
person: 80%
person: 46%
car: 67%
car: 99%
person: 50%
backpack: 32%
person: 30%
person: 71%
person: 88%
-----------------------------------------------------------------------------------------------
由于代码很多是yolov3的基础上修改的,因此原先在图片上增加置信值仍然可以使用,修改后就是这样,会挡住一些
这一张图片分辨率是800*485,在2G的940MX笔记本下,跑的时间是 Predicted in 214.287000 milli-seconds.