1. LabelImg標註文件製作VOC數據集
1.1 選取標註文件夾下的圖片與xml文件，並修改name（這一步其實有點多餘）
# -*- coding: UTF-8 -*-
import os
import shutil
# return xml_list
def find_xml(xml_fold):
xml_list = []
for _,dirs,files in os.walk(xml_fold):
for f in files:
if os.path.splitext(f)[-1]==".xml":
xml_list.append(f)
return xml_list
# return im_list
def find_image(im_fold):
im_list = []
im_ext_list = [".png",".PNG",".jpg",".JPG",".bmp",".BMP",".jpeg",".JPEG"]
for _,dirs,files in os.walk(im_fold):
for f in files:
if os.path.splitext(f)[-1] in im_ext_list:#'f' is an image?
im_list.append(f)
return im_list
def copy_xml_im(fold_path,xml_list,dst_fold):
xml_not_with_ext_list = []
for xml_with_ext in xml_list:
xml_not_with_ext_list.append(os.path.splitext(xml_with_ext)[-2])
# print(xml_not_with_ext_list)
im_ext_list = [".png",".PNG",".jpg",".JPG",".bmp",".BMP",".jpeg",".JPEG"]
for _,dirs,files in os.walk(fold_path):
for f in files:
# copy image
if os.path.splitext(f)[-1] in im_ext_list:#'f' is an image?
if os.path.splitext(f)[-2] in xml_not_with_ext_list:
shutil.copy(os.path.join(_,f),os.path.join(dst_fold,f))
# copy xml
if os.path.splitext(f)[-1] == ".xml":
shutil.copy(os.path.join(_,f),os.path.join(dst_fold,f))
def rename_xml_im(src_fold,dst_fold):
if os.path.exists(dst_fold)==False:
os.makedirs(dst_fold)
for _,dirs,files in os.walk(src_fold):
new_name = 0
for f in files:
if os.path.splitext(f)[-1]==".xml":
new_name = new_name + 1
shutil.copy(os.path.join(_,f),os.path.join(dst_fold,str(new_name).zfill(5)+".xml"))
shutil.copy(os.path.join(_,os.path.splitext(f)[-2]+".png"),os.path.join(dst_fold,str(new_name).zfill(5)+".png"))
def fix_xml_content(src_fold):
for _, dirs, files in os.walk(src_fold):
for f1 in files:
if (os.path.splitext(f1)[1]=='.xml'):
with open(os.path.join(src_fold,f1),'r') as f_r:
lines = f_r.readlines()
with open(os.path.join(src_fold, f1), 'w') as f_w:
for line in lines:
if '<filename>' in line:
line=' <filename>'+os.path.splitext(f1)[0]+'.png</filename>\n'
if '<path>' in line:
continue
f_w.write(line)
if __name__=="__main__":
fold_path = "/圖片與xml路徑"
dst_fold = fold_path + "/temp"
if os.path.exists(dst_fold)==False:
os.makedirs(dst_fold)
xml_list = find_xml(fold_path)
im_list = find_image(fold_path)
print("the number of xmlfile:",len(xml_list))
print("the number of image:",len(im_list))
copy_xml_im(fold_path,xml_list,dst_fold)
rename_xml_im(dst_fold,fold_path+"/rename")
fix_xml_content(fold_path+"/VOC2007")
1.2 將xml文件名隨機寫入train.txt/val.txt/test.txt
'''
find random index of [0,9] in every group.
select 6 numbers as train set index
select 2 numbers as val set index
select 2 numbers as test set index
number of groups = (number of xmlfile) / 10
others will be divided into test set
'''
import os
import random
def rand_index_list(start=0,end=9,proportion=1):
rand_index_list = []
while len(rand_index_list)<proportion*10:
rand_index=random.randint(start,end)
if rand_index not in rand_index_list:
rand_index_list.append(rand_index)
rand_index_list.sort()
return rand_index_list
def rand_serial(train_val_proportion,train_proportion):
rand_train_val_index_list=rand_index_list(proportion=train_val_proportion)
rand_train_index_list = [rand_train_val_index_list[x] for x in rand_index_list(start=0,end=train_proportion*10,proportion=train_proportion)]
return rand_train_val_index_list,rand_train_index_list
def find_rand_file(path_fold=None):
rand_train_val_index_list,rand_train_index_list = [],[]
xml_list = []
for _,dirs,files in os.walk(path_fold):
for f in files:
if os.path.splitext(f)[-1]==".xml":
xml_list.append(os.path.splitext(f)[-2])
# print(xml_list)
index_list_num = len(xml_list)/10
for index in range(index_list_num):
train_val_index_list,train_index_list = rand_serial(0.8,0.6)
rand_train_val_index_list = rand_train_val_index_list+[x+10*index for x in train_val_index_list]
rand_train_index_list = rand_train_index_list+[x+10*index for x in train_index_list]
rand_val_index_list = [x for x in rand_train_val_index_list if x not in rand_train_index_list]
# print("train_val_index:{}".format(rand_train_val_index_list))
# print("train_index:{}".format(rand_train_index_list))
# print("val_index:{}".format(rand_val_index_list))
train_list = [xml_list[i] for i in rand_train_index_list]
val_list = [xml_list[i] for i in rand_val_index_list]
test_list = [x for x in xml_list if (x not in train_list) and (x not in val_list)]
# print("train:{}".format(train_list))
# print("val:{}".format(val_list))
# print("test:{}".format(test_list))
return train_list,val_list,test_list
def write_file(voc_path):
train_list,val_list,test_list=find_rand_file(voc_path+"/Annotations")
with open(voc_path+"/ImageSets/Main/train.txt",'w') as f_w:
for x in train_list:
f_w.write(x)
f_w.write("\n")
with open(voc_path+"/ImageSets/Main/val.txt",'w') as f_w:
for x in val_list:
f_w.write(x)
f_w.write("\n")
with open(voc_path+"/ImageSets/Main/test.txt",'w') as f_w:
for x in test_list:
f_w.write(x)
f_w.write("\n")
if __name__ == "__main__":
write_file("VOC2007")
2. VOC數據集各類別面積大小分佈
2.1 說明
計算VOC數據集中各個類別（以我們自己的數據集爲例：'car',‘cottage','town house','apartment','person','bird nest','honeycomb'）的數量，以及各個類別的面積從0-16,32,64,128,256,512,1024，……的數量。並且繪製對應的直方圖
import os
from PIL import Image
import xml.etree.ElementTree as ET
import matplotlib.pyplot as plt
import numpy as np
def calculateObjectArea(xml, objectname, im=None, outputPath=None):
imWidth = 0
imHeight = 0
areatuple = []
tree = ET.parse(xml)
root = tree.getroot()
# print(root)
for child1 in root:
# print(child1.tag, child1.attrib, child1.text)
for child2 in child1:
if child2.tag == 'width':
imWidth = child2.text
# print(imWidth)
if child2.tag == 'height':
imHeight = child2.text
# print(imHeight)
if objectname in child2.text:
# print(child1)
for child3 in child1.iter('bndbox'):
for child4 in child3:
# print('child4:', child4.tag, child4.text)
if child4.tag == 'xmin':
xmin = child4.text
# print(type(child4.text))
if child4.tag == 'xmax':
xmax = child4.text
if child4.tag == 'ymin':
ymin = child4.text
if child4.tag == 'ymax':
ymax = child4.text
# print((int(xmax)-int(xmin))*(int(ymax)-int(ymin)))
areatuple.append((int(xmax)-int(xmin))*(int(ymax)-int(ymin)))
# print(areatuple)
areatuple.sort()
return areatuple
def drawHist(tup, objectname, width):
tup.sort()
x=tup
# print('x =',x)
# draw hist
n, bins, patches = plt.hist(x, width, normed = False, facecolor = 'g')
plt.xlabel('Object Area')
plt.ylabel('Number')
# add title
plt.title(objectname)
# add text
# plt.text(60, .025, r'$\mu=100,\ \sigma=15$')
# plt.axis([40, 160, 0, 0.03])
plt.grid(True, linewidth = '1')
# plt.show()
def searchxml(xmlpath, objectname):
sumtup,lower16,lower32,lower64=[],[],[],[]
lower128,lower256,lower384,lower512,lower640,lower768 = [],[],[],[],[],[]
lower1024,lower1152,lower1280,lower1536,lower1792=[],[],[],[],[]
lower2000,lower4000,lower6000=[],[],[]
count16,count32,count64 = 0,0,0
count128,count256 ,count384,count512,count640,count768 = 0,0,0,0,0,0
count1024,count1152,count1280,count1536,count1792=0,0,0,0,0
count2000,count4000,count6000 = 0,0,0
for _, dirs, files in os.walk(xmlpath):
for f1 in files:
if os.path.splitext(f1)[1] == '.xml' and '.idea' not in os.path.join(_, f1):
# print('file is ', f1)
temptup = calculateObjectArea(os.path.join(_, f1), objectname)
# print('temptup', temptup)
sumtup = sumtup + temptup
for x in sumtup:
if x > 0 and x < 16:
lower16.append(x)
count16 = count16 + 1
if x > 0 and x < 32:
lower32.append(x)
count32 = count32 + 1
if x > 0 and x < 64:
lower64.append(x)
count64 = count64 + 1
if x > 0 and x < 128:
lower128.append(x)
count128 = count128 + 1
if x > 0 and x < 256:
lower256.append(x)
count256 = count256 + 1
if x > 0 and x < 384:
lower384.append(x)
count384 = count384 + 1
if x > 0 and x < 512:
lower512.append(x)
count512 = count512 + 1
if x > 0 and x < 640:
lower640.append(x)
count640 = count640 + 1
if x > 0 and x < 768:
lower768.append(x)
count768 = count768 + 1
if x > 0 and x < 1024:
lower1024.append(x)
count1024 = count1024 + 1
if x > 0 and x < 1152:
lower1152.append(x)
count1152 = count1152 + 1
if x > 0 and x < 1280:
lower1280.append(x)
count1280 = count1280 + 1
if x > 0 and x < 1536:
lower1536.append(x)
count1536 = count1536 + 1
if x > 0 and x < 1792:
lower1792.append(x)
count1792 = count1792 + 1
if x > 0 and x < 2000:
lower2000.append(x)
count2000 = count2000 + 1
if x > 0 and x < 4000:
lower4000.append(x)
count4000 = count4000 + 1
if x > 0 and x < 6000:
lower6000.append(x)
count6000 = count6000 + 1
print(objectname+' total:', len(sumtup))
drawHist(sumtup, objectname +'\'s number:'+str(len(sumtup)), 200)
if count16:
print('the number of '+objectname+' (\'area<16\' pixel):', count16)
drawHist(lower16, objectname+'--the number of \'area<16\' pixel:'+str(count16), 30)
if count32:
print('the number of '+objectname+' (\'area<32\' pixel):', count32)
drawHist(lower32, objectname+'--the number of \'area<32\' pixel:'+str(count32), 30)
if count64:
print('the number of '+objectname+' (\'area<64\' pixel):', count64)
drawHist(lower64, objectname+'--the number of \'area<64\' pixel:'+str(count64), 30)
if count128:
print('the number of '+objectname+' (\'area<128\' pixel):', count128)
drawHist(lower128, objectname+'--the number of \'area<128\' pixel:'+str(count128), 30)
if count256:
print('the number of '+objectname+' (\'area<256\' pixel):', count256)
drawHist(lower256, objectname+'--the number of \'area<256\' pixel:'+str(count256), 30)
if count384:
print('the number of '+objectname+' (\'area<384\' pixel):', count384)
drawHist(lower384, objectname+'--the number of \'area<384\' pixel:'+str(count384), 30)
if count512:
print('the number of '+objectname+' (\'area<512\' pixel):', count512)
drawHist(lower512, objectname+'--the number of \'area<512\' pixel:'+str(count512), 30)
if count640:
print('the number of '+objectname+' (\'area<640\' pixel):', count640)
drawHist(lower640, objectname+'--the number of \'area<640\' pixel:'+str(count640), 30)
if count768:
print('the number of '+objectname+' (\'area<768\' pixel):', count768)
drawHist(lower768, objectname+'--the number of \'area<768\' pixel:'+str(count768), 50)
if count1024:
print('the number of '+objectname+' (\'area<1024\' pixel):', count1024)
drawHist(lower1024, objectname+'--the number of \'area<1024\' pixel:'+str(count1024), 50)
if count1152:
print('the number of '+objectname+' (\'area<1152\' pixel):', count1152)
drawHist(lower1152, objectname + '--the number of \'area<1152\' pixel:' + str(count1152), 60)
if count1280:
print('the number of '+objectname+' (\'area<1280\' pixel):', count1280)
drawHist(lower1280, objectname + '--the number of \'area<1280\' pixel:' + str(count1280), 80)
if count1536:
print('the number of '+objectname+' (\'area<1536\' pixel):', count1536)
drawHist(lower1536, objectname + '--the number of \'area<1536\' pixel:' + str(count1536), 80)
if count1792:
print('the number of '+objectname+' (\'area<1792\' pixel):', count1792)
drawHist(lower1792, objectname + '--the number of \'area<1792\' pixel:' + str(count1792), 100)
if count2000:
print('the number of '+objectname+' (\'area<2000\' pixel):', count2000)
drawHist(lower2000, objectname + '--the number of \'area<2000\' pixel:' + str(count2000), 100)
if count4000:
print('the number of '+objectname+' (\'area<4000\' pixel):', count4000)
drawHist(lower4000, objectname + '--the number of \'area<4000\' pixel:' + str(count4000), 100)
if count6000:
print('the number of '+objectname+' (\'area<6000\' pixel):', count6000)
drawHist(lower6000, objectname + '--the number of \'area<6000\' pixel:' + str(count6000), 100)
if __name__=='__main__':
xmlpath='./Annotations'
objectnameList = ['car','cottage','town house','apartment','person','bird nest','honeycomb']
for objectname in objectnameList:
searchxml(xmlpath, objectname)
重新上一個稍微好一點的代碼：
import os
import xml.etree.ElementTree as ET
def object_name_with_path_list(path_fold):
xml_with_path_list = []
for _,dirs,files in os.walk(path_fold):
for f in files:
if os.path.splitext(f)[-1]==".xml":
xml_with_path_list.append(os.path.join(_,f))
return xml_with_path_list
def get_object_area(path_fold):
xml_with_path_list = object_name_with_path_list(path_fold)
# print("xml_with_path:{}".format(xml_with_path_list))
area_list = []
for item in xml_with_path_list:
tree = ET.parse(item)
root = tree.getroot()
for child in root:
object_name_list = []
object_positon_list = []
if child.tag=="object":
for child1 in child:
if child1.tag == "name":
object_name = child1.text
object_name_list.append(object_name)
if child1.tag == "bndbox":
x_min = [child2.text for child2 in child1 if child2.tag=="xmin"]
x_max = [child2.text for child2 in child1 if child2.tag=="xmax"]
y_min = [child2.text for child2 in child1 if child2.tag=="ymin"]
y_max = [child2.text for child2 in child1 if child2.tag=="ymax"]
object_positon=x_min+y_min+x_max+y_max
object_positon_list.append(object_positon)
object_area = (int(x_max[0])-int(x_min[0]))*(int(y_max[0])-int(y_min[0]))
area_list.append(object_area)
# print("object_list:{0},{1}".format(object_name_list,object_positon_list))
area_lower_64_list = [x for x in area_list if x < 64]
area_lower_128_list = [x for x in area_list if x < 128]
area_lower_256_list = [x for x in area_list if x < 256]
area_lower_512_list = [x for x in area_list if x < 512]
print("total numbers:{},lower_64:{},lower_128:{},lower_256:{},lower_512:{}"\
.format(len(area_list),len(area_lower_64_list),len(area_lower_128_list),len(area_lower_256_list),len(area_lower_512_list)))
if __name__=="__main__":
get_object_area("VOC2007/Annotations/")
2.2 結果
直接上圖
3.裁剪labelImg標註的所有對象
# -*- coding:utf-8 -*-
'''
程序放在VOC2007文件夾下
'''
import os
import xml.etree.ElementTree as ET
from PIL import Image
import time
def mkdirs(outputPath):
if os.path.exists(outputPath) == False:
os.makedirs(outputPath)#make directory and subdirectory
# crop all object of VOC2007 datasets
def crop_object(xml_path,im_path):
name_list = [os.path.splitext(f)[-2] for _,dirs,files in os.walk(xml_path) for f in files if os.path.splitext(f)[-1]==".xml"]
for item in name_list:
tree = ET.parse(os.path.join(xml_path,item+".xml"))
root = tree.getroot()
object_name_list = [child.find('name').text for child in root if child.tag=="object"]
object_position_list = [(int(child.find('bndbox').find("xmin").text),
int(child.find('bndbox').find("ymin").text),
int(child.find('bndbox').find("xmax").text),
int(child.find('bndbox').find("ymax").text))
for child in root if child.tag=="object"]
im = Image.open(os.path.join(im_path, item+".jpg"))
im_after_crop = [im.crop(position) for position in object_position_list]
for crop_im_index in range(len(im_after_crop)):
object_name_path = object_name_list[crop_im_index]
now_time_stamp = int(time.time())
mkdirs(os.path.join("objects",object_name_path))
im_after_crop[crop_im_index].save(os.path.join("objects",object_name_path,str(now_time_stamp)+str(object_position_list[crop_im_index][0])+".jpg"))
if __name__ == '__main__':
crop_object("Annotations/","JPEGImages/")
4.選擇某一類對象object，刪除其餘所有對象
# -*- coding:utf-8 -*-
import os
import xml.etree.ElementTree as ET
import time
def mkdirs(outputPath):
if os.path.exists(outputPath) == False:
os.makedirs(outputPath)#make directory and subdirectory
# choose specific class(choose_class),remove all other classes.
def choose_object(xml_path,im_path,dst_path,choose_class):
name_list = [os.path.splitext(f)[-2] for _,dirs,files in os.walk(xml_path) for f in files if os.path.splitext(f)[-1]==".xml"]
for item in name_list:
tree = ET.parse(os.path.join(xml_path,item+".xml"))
root = tree.getroot()
object_item_list = [child for child in root if child.tag=="object" and child.find('name').text != choose_class]
object_cross_list = [child for child in root if child.tag=="object" and child.find('name').text == choose_class]
for object_item in object_item_list:
root.remove(object_item)
mkdirs(dst_path)
if len(object_cross_list)>0:
tree.write(os.path.join(dst_path,item+".xml"))
if __name__ == '__main__':
choose_object("Annotations/","JPEGImages/","dst_xml_path/","car")