arnoldfychen / Yolo-Fastest

:zap: Yolo universal target detection model combined with EfficientNet-lite, the calculation amount is only 230Mflops(0.23Bflops), and the model size is 1.3MB

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

image

Yolo-Fastest

  • Simple, fast, compact, easy to transplant
  • A real-time target detection algorithm for all platforms
  • The fastest and smallest known universal target detection algorithm based on yolo
  • Optimized design for ARM mobile terminal, optimized to support NCNN reasoning framework
  • The speed is 45% faster than mobilenetv2-yolov3-nano, and the parameter amount is reduced by 56%

Evaluating indicator

Network VOC mAP(0.5) Resolution Run Time(Ncnn 1xCore) Run Time(Ncnn 4xCore) FLOPS Weight size
MobileNetV2-YOLOv3-Nano 65.27 320 11.36ms 5.48ms 0.55BFlops 3.0MB
Yolo-Fastest(our) 61.02 320 6.74ms 4.42ms 0.23BFlops 1.3MB
MobileNetV2 SSD-Lite 68.6 300 &ms &ms &BFlops 13.8MB
Yolo-Fastest-XL(our) 68.8 320 15.15ms 7.09ms 0.70BFlops 3.5MB
  • Test platform Kirin 990,Based on NCNN
  • Suitable for hardware with extremely tight computing resources
  • This model is recommended to do some simple single object detection suitable for simple application scenarios

Test

image

How to Train

Generate a pre-trained model for the initialization of the model backbone

  ./darknet partial yolo-fastest.cfg yolo-fastest.weights yolo-fastest.conv.109 109

Train

  ./darknet detector train voc.data yolo-fastest.cfg yolo-fastest.conv.109 

Thanks

About

:zap: Yolo universal target detection model combined with EfficientNet-lite, the calculation amount is only 230Mflops(0.23Bflops), and the model size is 1.3MB

License:Other


Languages

Language:C 63.9%Language:Cuda 15.1%Language:C++ 13.2%Language:Python 4.5%Language:CMake 1.4%Language:Batchfile 0.6%Language:Shell 0.5%Language:PowerShell 0.4%Language:Makefile 0.3%Language:C# 0.1%