darknet

README

go-darknet: Go bindings for Darknet (Yolo V4, Yolo V3)

go-darknet is a Go package, which uses Cgo to enable Go applications to use YOLO V4/V3 in Darknet.

Since this repository https://github.com/gyonluks/go-darknet is no longer maintained I decided to move on and make little different bindings for Darknet.

This bindings aren't for official implementation but for AlexeyAB's fork.

Paper Yolo v4: https://arxiv.org/abs/2004.10934

Paper Yolo v3: https://arxiv.org/abs/1804.02767

Table of Contents

• Why
• Requirements
• Installation
• Usage
• Documentation
• License

Why

Why does this repository exist?

Because this repository https://github.com/gyonluks/go-darknet is no longer maintained.

What is purpose of this bindings when you can have GoCV (bindings to OpenCV) and it handle Darknet YOLO perfectly?

Well, you don't need bunch of OpenCV dependencies and OpenCV itself sometimes.

Example of such project here: https://github.com/LdDl/license_plate_recognition#license-plate-recognition-with-go-darknet---- .

Requirements

You need to install fork of darknet. Latest commit I've tested is here. It corresponds last official YOLOv4 release

Use provided Makefile.

• For CPU-based instalattion:

  make install_darknet
  

• For both CPU and GPU-based instalattion if you HAVE CUDA installed:

  make install_darknet_gpu
  

  Note: I've tested CUDA 10.2 and cuDNN is 7.6.5)

• For both CPU and GPU-based instalattion if you HAVE NOT CUDA installed:

  make install_darknet_gpu_cuda
  

  Note: There is some struggle in Makefile for cuDNN, but I hope it works in Ubuntu atleast. Do not forget provide proper CUDA and cuDNN versions.

Installation

go get github.com/LdDl/go-darknet

Usage

Example Go program is provided in the examples directory. Please refer to the code on how to use this Go package.

Building and running program:

• Navigate to examples folder

  cd ${YOUR PATH}/github.com/LdDl/go-darknet/cmd/examples
  

• Download dataset (sample of image, coco.names, yolov4.cfg (or v3), yolov4.weights(or v3)).

  #for yolo v4
  ./download_data.sh
  #for yolo v3
  ./download_data_v3.sh
  

• Note: you don't need coco.data file anymore, because sh-script above does insert coco.names into 'names' field in yolov4.cfg file (so AlexeyAB's fork can deal with it properly) So last rows in yolov4.cfg file will look like:

  ......
  [yolo]
  .....
  iou_loss=ciou
  nms_kind=greedynms
  beta_nms=0.6
  
  names = coco.names # this is path to coco.names file
  

• Also do not forget change batch and subdivisions sizes from:

  batch=64
  subdivisions=8
  

  to

  batch=1
  subdivisions=1
  

  It will reduce amount of VRAM used for detector test.

• Build and run example program Yolo V4:

  go build -o base_example/main base_example/main.go && ./base_example/main --configFile=yolov4.cfg --weightsFile=yolov4.weights --imageFile=sample.jpg
  

  Output should be something like this:

  traffic light (9): 73.5040% | start point: (238,73) | end point: (251, 106)
  truck (7): 96.6401% | start point: (95,79) | end point: (233, 287)
  truck (7): 96.4774% | start point: (662,158) | end point: (800, 321)
  truck (7): 96.1841% | start point: (0,77) | end point: (86, 333)
  truck (7): 46.8694% | start point: (434,173) | end point: (559, 216)
  car (2): 99.7370% | start point: (512,188) | end point: (741, 329)
  car (2): 99.2532% | start point: (260,191) | end point: (422, 322)
  car (2): 99.0333% | start point: (425,201) | end point: (547, 309)
  car (2): 83.3920% | start point: (386,210) | end point: (437, 287)
  car (2): 75.8621% | start point: (73,199) | end point: (102, 274)
  car (2): 39.1925% | start point: (386,206) | end point: (442, 240)
  bicycle (1): 76.3121% | start point: (189,298) | end point: (253, 402)
  person (0): 97.7213% | start point: (141,129) | end point: (283, 362)
  

  Yolo V3:

  go build main.go && ./main --configFile=yolov3.cfg --weightsFile=yolov3.weights --imageFile=sample.jpg
  

  Output should be something like this:

  truck (7): 49.5123% | start point: (0,136) | end point: (85, 311)
  car (2): 36.3694% | start point: (95,152) | end point: (186, 283)
  truck (7): 48.4177% | start point: (95,152) | end point: (186, 283)
  truck (7): 45.6520% | start point: (694,178) | end point: (798, 310)
  car (2): 76.8402% | start point: (1,145) | end point: (84, 324)
  truck (7): 25.5920% | start point: (107,89) | end point: (215, 263)
  car (2): 99.8782% | start point: (511,185) | end point: (748, 328)
  car (2): 99.8193% | start point: (261,189) | end point: (427, 322)
  car (2): 99.6405% | start point: (426,197) | end point: (539, 311)
  car (2): 74.5627% | start point: (692,186) | end point: (796, 316)
  car (2): 72.7975% | start point: (388,206) | end point: (437, 276)
  bicycle (1): 72.2760% | start point: (178,270) | end point: (268, 406)
  person (0): 97.3007% | start point: (143,135) | end point: (268, 343)
  

Documentation

See go-darknet's API documentation at GoDoc.

License

go-darknet follows Darknet's license.

Documentation

Index

• type BoundingBox
• type DarknetImage
  • func Float32ToDarknetImage(flatten []float32, width, height int) (*DarknetImage, error)
  • func Image2Float32(src image.Image) (*DarknetImage, error)
  • func ImageRGBA2Float32(img *image.RGBA) (*DarknetImage, error)
  • func (img *DarknetImage) Close() error
• type Detection
• type DetectionResult
• type YOLONetwork
  • func (n *YOLONetwork) Close() error
  • func (n *YOLONetwork) Detect(img *DarknetImage) (*DetectionResult, error)
  • func (n *YOLONetwork) Init() error

Types

type BoundingBox

type BoundingBox struct {
	StartPoint image.Point
	EndPoint   image.Point
}

BoundingBox represents a bounding box.

type DarknetImage

type DarknetImage struct {
	Width  int
	Height int
	// contains filtered or unexported fields
}

DarknetImage represents the image buffer.

func Float32ToDarknetImage

func Float32ToDarknetImage(flatten []float32, width, height int) (*DarknetImage, error)

Float32ToDarknetImage Converts []float32 to darknet image

func Image2Float32

func Image2Float32(src image.Image) (*DarknetImage, error)

func ImageRGBA2Float32

func ImageRGBA2Float32(img *image.RGBA) (*DarknetImage, error)

Image2Float32 Returns []float32 representation of image.Image

func (*DarknetImage) Close

func (img *DarknetImage) Close() error

Close and release resources.

type Detection

type Detection struct {
	BoundingBox

	ClassIDs      []int
	ClassNames    []string
	Probabilities []float32
}

Detection represents a detection.

type DetectionResult

type DetectionResult struct {
	Detections           []*Detection
	NetworkOnlyTimeTaken time.Duration
	OverallTimeTaken     time.Duration
}

DetectionResult represents the inference results from the network.

type YOLONetwork

type YOLONetwork struct {
	GPUDeviceIndex           int
	NetworkConfigurationFile string
	WeightsFile              string
	Threshold                float32

	ClassNames []string
	Classes    int
	// contains filtered or unexported fields
}

YOLONetwork represents a neural network using YOLO.

func (*YOLONetwork) Close

func (n *YOLONetwork) Close() error

Close and release resources.

func (*YOLONetwork) Detect

func (n *YOLONetwork) Detect(img *DarknetImage) (*DetectionResult, error)

Detect specified image

func (*YOLONetwork) Init

func (n *YOLONetwork) Init() error

Init the network.