tchigher / 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

• 中文介绍https://zhuanlan.zhihu.com/p/234506503

• 9.12：更新NCNN Camera Demo https://github.com/dog-qiuqiu/Yolo-Fastest/tree/master/ncnn_sample

• On some GPUs (such as NVIDIA PASCAL: 1080ti, 1070...), Darknet Group convolution is not well supported, which will cause the problem of low training inference efficiency, but it will not appear on the 20 series and 16 series graphics cards, for example The reasoning time for 2080ti is 2ms, and 1660ti is 3ms. It is suspected to be the cause of CUDNN. It is recommended that users in this situation use pytorch for training and inference

• Based on pytorch training framework: https://github.com/dog-qiuqiu/yolov3

⚡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
• Based on NCNN deployed on RK3399 ,Raspberry Pi 4b... and other embedded devices to achieve full real-time 30fps+
• The speed is 45% faster than mobilenetv2-yolov3-nano, and the parameter amount is reduced by 56%

Evaluating indicator/Benchmark

Network	VOC mAP(0.5)	COCO mAP(0.5)	Resolution	Run Time(Ncnn 1xCore)	Run Time(Ncnn 4xCore)	FLOPS	Weight size
MobileNetV2-YOLOv3-Nano	65.27	30.13	320	11.36ms	5.48ms	0.55BFlops	3.0MB
Yolo-Fastest(our)	61.02	&	320	6.74ms	4.42ms	0.23BFlops	1.3MB
Yolo-Fastest-XL(our)	69.43	32.45	320	15.15ms	7.09ms	0.70BFlops	3.5MB

• Test platform Kirin 990 CPU，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

Pascal VOC performance index comparison

Network	Model Size	mAP(VOC 2017)	FLOPS
Tiny YOLOv2	60.5MB	57.1%	6.97BFlops
Tiny YOLOv3	33.4MB	58.4%	5.52BFlops
YOLO Nano	4.0MB	69.1%	4.51Bflops
MobileNetv2-SSD-Lite	13.8MB	68.6%	&Bflops
MobileNetV2-YOLOv3	11.52MB	70.20%	2.02Bflos
Pelee-SSD	21.68MB	70.09%	2.40Bflos
Yolo Fastest	1.3MB	61.02%	0.23Bflops
Yolo Fastest-XL	3.5MB	69.43%	0.70Bflops
MobileNetv2-Yolo-Lite	8.0MB	73.26%	1.80Bflops

• Performance indicators reference from the papers and public indicators in the github project
• MobileNetv2-Yolo-Lite: https://github.com/dog-qiuqiu/MobileNet-Yolo#mobilenetv2-yolov3-litenano-darknet

Raspberrypi 3b Ncnn bf16s benchmark(4xA53 1.2Ghz)

loop_count = 4
num_threads = 4
powersave = 0
gpu_device = -1
cooling_down = 1
        yolo-fastest  min =   62.58  max =   62.76  avg =   62.70
      squeezenet_ssd  min =  380.98  max =  391.39  avg =  387.53
 squeezenet_ssd_int8  min =  458.05  max =  467.54  avg =  463.12
       mobilenet_ssd  min =  212.31  max =  223.34  avg =  218.93
  mobilenet_ssd_int8  min =  359.98  max =  374.03  avg =  365.17
      mobilenet_yolo  min =  619.65  max =  635.44  avg =  628.29
  mobilenetv2_yolov3  min =  294.92  max =  304.95  avg =  298.43
         yolov4-tiny  min =  855.50  max = 1074.92  avg =  962.78

Compile

How to compile on Linux

• This repo is based on Darknet project so the instructions for compiling the project are same (https://github.com/MuhammadAsadJaved/darknet#how-to-compile-on-windows-legacy-way)

Just do make in the Yolo-Fastest-master directory. Before make, you can set such options in the Makefile: link

• GPU=1 to build with CUDA to accelerate by using GPU (CUDA should be in /usr/local/cuda)
• CUDNN=1 to build with cuDNN v5-v7 to accelerate training by using GPU (cuDNN should be in /usr/local/cudnn)
• CUDNN_HALF=1 to build for Tensor Cores (on Titan V / Tesla V100 / DGX-2 and later) speedup Detection 3x, Training 2x
• OPENCV=1 to build with OpenCV 4.x/3.x/2.4.x - allows to detect on video files and video streams from network cameras or web-cams
• Set the other options in the Makefile according to your need.

Test/Demo

*Run Yolo-Fastest , Yolo-Fastest-xl , Yolov3 or Yolov4 on image or video inputs

Demo on image input

*Note: change .data , .cfg , .weights and input image file in image_yolov3.sh for Yolo-Fastest-x1, Yolov3 and Yolov4

  sh image_yolov3.sh

Demo on video input

*Note: Use any input video and place in the data folder or use 0 in the video_yolov3.sh for webcam

*Note: change .data , .cfg , .weights and input video file in video_yolov3.sh for Yolo-Fastest-x1, Yolov3 and Yolov4

  sh video_yolov3.sh

Yolo-Fastest Test

Yolo-Fastest-xl Test

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

• 交流qq群:1062122604
• https://github.com/AlexeyAB/darknet

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

Deploy

NCNN

NCNN Conversion Tutorial

• Benchmark:https://github.com/Tencent/ncnn/tree/master/benchmark
• NCNN supports direct conversion of darknet models
• darknet2ncnn: https://github.com/Tencent/ncnn/tree/master/tools/darknet

NCNN C++ Sample

• https://github.com/dog-qiuqiu/Yolo-Fastest/tree/master/ncnn_sample

NCNN Android Sample

• https://github.com/WZTENG/YOLOv5_NCNN

MNN&TNN&MNN

• https://github.com/dog-qiuqiu/MobileNet-Yolo#darknet2caffe-tutorial

ONNX&TensorRT

• https://github.com/CaoWGG/TensorRT-YOLOv4
• It is not efficient to run on Psacal and earlier GPU architectures. It is not recommended to deploy on such devices such as jeston nano(17ms/img), Tx1, Tx2, but there is no such problem in Turing GPU, such as jetson-Xavier-NX Can run efficiently

Thanks

• https://github.com/AlexeyAB/darknet
• https://github.com/Tencent/ncnn