• 2021.8.12: yolo-fastestV2版本已发布:https://github.com/dog-qiuqiu/Yolo-FastestV2
• 2021.3.21: 对模型结构进行细微调整优化，更新Yolo-Fastest-1.1模型
• 2021.3.19: NCNN Camera Demo https://github.com/dog-qiuqiu/Yolo-Fastest/tree/master/sample/ncnn
• 2021.3.16: 修复分组卷积在某些旧架构GPU推理耗时异常的问题

⚡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+

• 中文介绍https://zhuanlan.zhihu.com/p/234506503
• 相比AlexeyAB/darknet，此版本的darknet修复分组卷积在某些旧架构GPU推理耗时异常的问题(例如1050ti:40ms->4ms速度提升10倍)，强烈建议用此仓库框架训练模型
• Compared with AlexeyAB/darknet, this version of darknet fixes the problem of abnormal time-consuming inference of grouped convolution in some old architecture GPUs (for example, 1050ti:40ms->4ms speed up 10 times), it is strongly recommended to use this warehouse framework for training model
• Darknet CPU推理效率优化不好，不建议使用Darknet作为CPU端的推理框架，建议使用NCNN
• Darknet CPU reasoning efficiency optimization is not good, it is not recommended to use Darknet as the CPU side reasoning framework, it is recommended to use ncnn
• Based on pytorch training framework: https://github.com/dog-qiuqiu/yolov3

Evaluating indicator/Benchmark

• Test platform Mi 11 Snapdragon 888 CPU，Based on NCNN
• COCO 2017 Val mAP（no group label）
• Suitable for hardware with extremely tight computing resources
• This model is recommended to do some simple single object detection suitable for simple application scenarios

Yolo-Fastest-1.1 Multi-platform benchmark

• The above is a multi-core test benchmark
• The above speed benchmark is tested by big core in big.little CPU
• Raspberrypi 3B enable bf16s optimization，Raspberrypi 64 Bit OS
• Rk3399 needs to lock the cpu to the highest frequency, ncnn and enable bf16s optimization

Pascal VOC performance index comparison

• 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

Yolo-Fastest-1.1 Pedestrian detection

• Simple real-time pedestrian detection model based on yolo-fastest-1.1
• Enable bf16s optimization，Raspberrypi 64 Bit OS

Demo

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 Sample

• CamSample:https://github.com/dog-qiuqiu/Yolo-Fastest/tree/master/sample/ncnn
• AndroidSample: https://github.com/WZTENG/YOLOv5_NCNN

MNN&TNN&MNN

• https://github.com/dog-qiuqiu/MobileNet-Yolo#darknet2caffe-tutorial
• Based on MNN: https://github.com/geekzhu001/Yolo-Fastest-MNN Run on : raspberry pi 4B 2G Input size : 320320 Average inference time : 0.035s*

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

OpenCV DNN

• https://blog.csdn.net/nihate/article/details/108670542

Thanks

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

Cite as

dog-qiuqiu. (2021, July 24). dog-qiuqiu/Yolo-Fastest: yolo-fastest-v1.1.0 (Version v.1.1.0). Zenodo. http://doi.org/10.5281/zenodo.5131532