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Since this repository https://github.com/gyonluks/go-darknet is no longer maintained I decided to move on and make little different bindings for Darknet.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
Requirements
For proper codebase please use fork of darknet. Latest commit I've tested here
In order to use go-darknet, libdarknet.so should be available in one of
the following locations:
- /usr/lib
- /usr/local/lib
Also, darknet.h should be available in one of the following locations:
- /usr/include
- /usr/local/include
To achieve it, after Darknet compilation (via make) execute following command:
# Copy *.so to /usr/lib + /usr/include (or /usr/local/lib + /usr/local/include)
sudo cp libdarknet.so /usr/lib/libdarknet.so && sudo cp include/darknet.h /usr/include/darknet.h
# sudo cp libdarknet.so /usr/local/lib/libdarknet.so && sudo cp include/darknet.h /usr/local/include/darknet.h
Note: do not forget to set LIBSO=1 in Makefile before executing 'make':
LIBSO=1
Installation
go get github.com/LdDl/go-darknet
Usage
Example Go program is provided in the example directory. Please refer to the code on how to use this Go package.
Building and running program:
Navigate to example folder
cd $GOPATH/github.com/LdDl/go-darknet/example
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 program Yolo V4:
go build main.go && ./main --configFile=yolov4.cfg --weightsFile=yolov4.weights --imageFile=sample.jpg
Output should be something like this:
traffic light (9): 73.5039% | 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): 46.8695% | start point: (434,173) | end point: (559, 216)
car (2): 99.7370% | start point: (512,188) | end point: (741, 329)
car (2): 99.2533% | start point: (260,191) | end point: (422, 322)
car (2): 99.0333% | start point: (425,201) | end point: (547, 309)
car (2): 83.3919% | 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): 49.5197% | start point: (0,136) | end point: (85, 311)
car (2): 36.3747% | start point: (95,152) | end point: (186, 283)
truck (7): 48.4384% | start point: (95,152) | end point: (186, 283)
truck (7): 45.6590% | start point: (694,178) | end point: (798, 310)
car (2): 76.8379% | start point: (1,145) | end point: (84, 324)
truck (7): 25.5731% | start point: (107,89) | end point: (215, 263)
car (2): 99.8783% | start point: (511,185) | end point: (748, 328)
car (2): 99.8194% | start point: (261,189) | end point: (427, 322)
car (2): 99.6408% | start point: (426,197) | end point: (539, 311)
car (2): 74.5610% | start point: (692,186) | end point: (796, 316)
car (2): 72.8053% | start point: (388,206) | end point: (437, 276)
bicycle (1): 72.2932% | start point: (178,270) | end point: (268, 406)
person (0): 97.3026% | start point: (143,135) | end point: (268, 343)
Documentation
See go-darknet's API documentation at GoDoc.