ft 416

【⽬标检测】Yolov5训练Crowdhuman数据集

这篇⽂章的⽬的，1.解析Crowdhuman数据集；2.准备Crowdhuman相应配置，训练yolov5。

其中我的是将273271,，全部写为273271_

1.环境

ubuntu16.04

cuda10.1

cudnn7

python3.6

Cython

matplotlib>=3.2.2

numpy>=1.18.5

opencv-python>=4.1.2

Pillow

PyYAML>=5.3

scipy>=1.4.1

tensorboard>=2.2

torch>=1.7.0(my1.7.1)

torchvision>=0.8.1(my0.8.2)

tqdm>=4.41.0

aborn>=0.11.0

easydict

thop

pycocotools>=2.0

2.下载

（1）crowdhuman数据集：

（2）yolov5的⼯程

（3）yolov5预训练模型下载

放到yolov5/weights⽬录下

3.解析Crowdhuman数据集

这⾥有两个参考：参考的处理crowdhuman数据集的源码，另⼀个是⼀篇应⽤的博客。

由于我要的图⽚格式和作者的稍微有点不同，但是还是需要花点时间调试⼀下。⽐如⽬录结构，以及odgt中图⽚名称等。下⾯进⼊正题。

（1）⾸先将下载的源码YOLOv5-Tools/CrowHuman2YOLO/data⽬录下的数据，放到yolov5/data⽬录下：

（2）将下载完成的数据，放到raw⽬录下

（3）修改脚本和

#prepare_47-48inrtthis:

cd../crowdhuman-$1/

forfilein`ls|`

do

newfile=`echo$file|d's/,/_/g'`

mv$file$newfile

echoRenamefile:$newfile...

done

#gen_geline80:

ID=anno['ID']#e.g.'273271,c9db000d5146c15'

#to

ID=anno['ID'].replace(',','_')#e.g.'273271,c9db000d5146c15'

完整代码和：

#!/bin/bash

#prepare_

t-e

#checkargument

if[[-z$1||!$1=~[[:digit:]]x[[:digit:]]]];then

echo"ERROR:Thisscriptrequires1argument,"inputdimension"oftheYOLOmodel."

echo"Theinputdimensionshouldbe{width}x{height}suchas608x608or416x256.".

exit1

fi

ifwhichpython3>/dev/null;then

PYTHON=python3

el

PYTHON=python

fi

pushd$(dirname$0)/raw>/dev/null

get_file()

{

#dodownloadonlyifthefiledoesnotexist

if[[-f$2]];then

echoSkipping$2

el

echoDownloading$2...

$1

fi

}

echo"**Downloaddatatfiles"

#unzipimagefiles(ignoreCrowdHuman_now)

echo"**Unzipdatatfiles"

forfinCrowdHuman_wdHuman_wdHuman_wdHuman_;do

unzip-n${f}

done

echo"**Createthecrowdhuman-$1/subdirectory"

rm-rf../crowdhuman-$1/

mkdir../crowdhuman-$1/

lnImages/*.jpg../crowdhuman-$1/

#thecrowdhuman/subdirectorynowcontainsalltrain/valjpgimages

#changefile','with'_'

cd../crowdhuman-$1/

forfilein`ls|`

do

newfile=`echo$file|d's/,/_/g'`

mv$file$newfile

echoRenamefile:$newfile...

done

echo"**Generateyolotxtfiles"

cd..

${PYTHON}gen_$1

popd>/dev/null

echo"**Done."

"""gen_

TogenerateYOLOtxtfilesfromtheoriginalCrowdHumanannotations.

isdirectory.

Inputs:

*raw/annotation_

*raw/annotation_

*crowdhuman-{width}x{height}/[IDs].jpg

Outputs:

*crowdhuman-{width}x{height}

*crowdhuman-{width}x{height}/

*crowdhuman-{width}x{height}/[IDs].txt(oneannotationforeachimageinthetrainingortestt)

"""

importjson

frompathlibimportPath

fromargparimportArgumentParr

fromargparimportArgumentParr

importnumpyasnp

importcv2

#inputimagewidth/heightoftheyolov4model,tbycommand-lineargument

INPUT_WIDTH=0

INPUT_HEIGHT=0

#Minimumwidth/heightofobjectsfordetection(don'tlearnfrom

#objectssmallerthanthe

MIN_W=5

MIN_H=5

#DoK-Meansclusteringinordertodetermine"anchor"sizes

DO_KMEANS=True

KMEANS_CLUSTERS=9

BBOX_WHS=[]#keeptrackofbboxwidth/heightwithrespectto608x608

defimage_shape(ID,image_dir):

asrtimage_dirisnotNone

jpg_path=image_dir/('%'%ID)

img=(jpg__posix())

deftxt_line(cls,bbox,img_w,img_h):

"""Generate1lineinthetxtfile."""

asrtINPUT_WIDTH>0andINPUT_HEIGHT>0

x,y,w,h=bbox

x=max(int(x),0)

y=max(int(y),0)

w=min(int(w),img_w-x)

h=min(int(h),img_h-y)

w_rescaled=float(w)*INPUT_WIDTH/img_w

h_rescaled=float(h)*INPUT_HEIGHT/img_h

ifw_rescaled

return''

el:

ifDO_KMEANS:

globalBBOX_WHS

BBOX_((w_rescaled,h_rescaled))

cx=(x+w/2.)/img_w

cy=(y+h/2.)/img_h

nw=float(w)/img_w

nh=float(h)/img_h

return'%d%.6f%.6f%.6f%.6fn'%(cls,cx,cy,nw,nh)

defprocess(t_='test',annotation_filename='raw/annotation_',

output_dir=None):

"""Procesither'train'or'test't."""

asrtoutput_dirisnotNone

output_(exist_ok=True)

jpgs=[]

withopen(annotation_filename,'r')asfanno:

forraw_nes():

anno=(raw_anno)

ID=anno['ID'].raplace(',','_')#e.g.'273271_c9db000d5146c15'

print('ProcessingID:%s'%ID)

img_h,img_w,img_c=image_shape(ID,output_dir)

asrtimg_c==3#shouldbeaBGRimage

txt_path=output_dir/('%'%ID)

#writeatxtforeachimage

#writeatxtforeachimage

withopen(txt__posix(),'w')asftxt:

forobjinanno['gtboxes']:

ifobj['tag']=='mask':

continue#ignorenon-human

asrtobj['tag']=='person'

if'hbox'():#head

line=txt_line(0,obj['hbox'],img_w,img_h)

ifline:

(line)

if'fbox'():#fullbody

line=txt_line(1,obj['fbox'],img_w,img_h)

ifline:

(line)

('data/%s/%'%(output_dir,ID))

#writethe'data/crowdhuman/'or'data/crowdhuman/'

t_path=output_dir/('%'%t_)

withopen(t__posix(),'w')asft:

forjpginjpgs:

('%sn'%jpg)

defrm_txts(output_dir):

"""Removetxtfilesinoutput_dir."""

fortxtinoutput_('*.txt'):

_file():

()

defmain():

globalINPUT_WIDTH,INPUT_HEIGHT

parr=ArgumentParr()

_argument('dim',help='inputwidthandheight,e.g.608x608')

args=_args()

dim_split=('x')

iflen(dim_split)!=2:

raiSystemExit('ERROR:badspecofinputdim(%s)'%)

INPUT_WIDTH,INPUT_HEIGHT=int(dim_split[0]),int(dim_split[1])

ifINPUT_WIDTH%32!=0orINPUT_HEIGHT%32!=0:

raiSystemExit('ERROR:badspecofinputdim(%s)'%)

output_dir=Path('crowdhuman-%s'%)

ifnotoutput__dir():

raiSystemExit('ERROR:%sdoesnotexist.'%output__posix())

rm_txts(output_dir)

process('test','raw/annotation_',output_dir)

process('train','raw/annotation_',output_dir)

withopen('crowdhuman-%'%,'w')asf:

("""class=2

train=data/crowdhuman-%s/

valid=data/crowdhuman-%s/

names=data/

backup=backup/n"""%(,))

ifDO_KMEANS:

try:

rimportKMeans

exceptModuleNotFoundError:

print('WARNING:nosklearn,skippinganchorclustering...')

el:

X=(BBOX_WHS)

kmeans=KMeans(n_clusters=KMEANS_CLUSTERS,random_state=0).fit(X)

kmeans=KMeans(n_clusters=KMEANS_CLUSTERS,random_state=0).fit(X)

centers=r_centers_

centers=centers[centers[:,0].argsort()]#sortbybboxw

print('n**foryolov5-%dx%d,'%(INPUT_WIDTH,INPUT_HEIGHT),end='')

print('resizedbboxwidth/heightclustersare:',end='')

print(''.join(['(%.2f,%.2f)'%(c[0],c[1])forcincenters]))

print('nanchors=',end='')

print(','.join(['%d,%d'%(int(c[0]),int(c[1]))forcincenters]))

if__name__=='__main__':

main()

（4）运⾏数据处理脚本

bash./prepare_608x608

4.训练

添加⼀个yaml配置⽂件在data⽬录下：

#

train:"./data/crowdhuman-608x608/"

val:"./data/crowdhuman-608x608/"

nc:2

names:["head","person"]

进⼊yolov5⽬录下，运⾏以下脚本（其中models中选择的模型⼤⼩配置⽂件，与weights中的预训练模型应该是⼀致）：

--datadata/--cfgmodels/--weightsweights/--device0

这⾥就欧克了。