YOLOv4-Tiny来了!371 FPS!
前言
今天刷屏的动态一定是 YOLOv4-Tiny!
实际上,YOLOv4-Tiny 在大前天(2020.06.25)的晚上就正式发布了,但鉴于当时处于端午假期,Amusi 特意没有更新,希望各位CVers过个好节,科研缓一缓,哈哈。
YOLOv4 重要时间节点 2020.04:YOLOv4 正式发布 2020.06:YOLOv4-Tiny 正式发布
YOLOv4-Tiny来了!
时隔两个月,YOLOv4-Tiny版本正式推出! 在COCO上的性能:40.2% AP50, 371 FPS (GTX 1080 Ti)
无论是AP,还是FPS的性能,其相较于YOLOv3-Tiny、Pelee、CSP都是巨大的提升,如下图所示:
目前YOLOv4的模型(cfg、weights)已经在官网放出,大家可以下载(我也已经打包好,详见文末),模型仅占23.1 MB! https://github.com/AlexeyAB/darknet
很多同学一定会问,YOLOv4-Tiny相较于YOLOv4有哪些变化?删减了哪些模块或者layers?
不过,Amusi 也调查到了关于YOLOv4-Tiny的细节,AlexeyAB 在github回复网友提问的内容如下:
-
There is used resize=1.5 instead of random=1, that you suggested, congrats! #3830 CSP: There is used groups for [route] layer for CSP - EFM: http://openaccess.thecvf.com/content_CVPRW_2020/papers/w28/Wang_CSPNet_A_New_Backbone_That_Can_Enhance_Learning_Capability_of_CVPRW_2020_paper.pdf Initially it was done for MixNet: #4203 (comment) There is used CIoU-loss with optimal normalizers (as in YOLOv4) scale_x_y parameter (as in YOLOv4)
想了解更多YOLOv4-Tiny内容,其实建议上github,直接和AlexeyAB对线交流,他回复的相当及时,链接如下,推荐直接交流: https://github.com/AlexeyAB/darknet/issues/6067
从上面的回复,我才知道CSPNet收录到了CVPR 2020 Workshop,个人非常喜欢这篇论文,不管是YOLOv4还是后面所谓的"YOLOv5",都用到了CSPNet,贡献涨点很明显!推荐精读这篇论文: http://openaccess.thecvf.com/content_CVPRW_2020/papers/w28/Wang_CSPNet_A_New_Backbone_That_Can_Enhance_Learning_Capability_of_CVPRW_2020_paper.pdf
YOLOv4-Tiny测试
国外已经有网友亲测了YOLOv4-Tiny,并在Titan XP、Jetson AGX Xavier两个设备上测试,结果如下:
Titan XP
Jetson AGX Xavier
速度真的相当快!AlexeyAB 还透露到将来OpenCV或者TensorRT版本的YOLOv4-Tiny速度会高达 500 - 1000 FPS!
@CSTEZCAN Thanks! I think yolov4-tiny can work with 500 - 1000 FPS by using OpenCV or tkDNN/TensorRT when it will be implemented in these libraries.
侃侃
Doesn’t matter what I think! At this point @alexeyab84 has the canonical version of darknet and yolo, he’s put a ton of work into it and everyone uses it, not mine haha.
而且YOLOv1-v3的官方项目已经把YOLOv4的项目链接挂上去了。
关于YOLOv4,详见:
彩蛋
YOLOv4-Tiny模型下载
我已经将模型、论文等资源打包好,网络访问受限的同学可以使用这个百度云链接:
链接:https://pan.baidu.com/s/1kvrh4hh_aZj7LZFTRULLZQ 提取码:bocz
