写这篇主要是写了代码不想浪费
用的是github上的tf-faster-rcnn（自己搜下）
python3.6.2版本
错误：assert（boxes[:,2]>=boxes[:,0]）.all():
boxes是标注的bounding box,检查Xmax是否大于Xmin,只要有一个不符合就报错
======================
问题原因：1.数据集有问题2.数据集改过来以后没删除缓存
2问题看https://blog.csdn.net/10km/article/details/64641322，他用的是github上的py-faster-rcnn如果你用的跟我一样，是github上的tf-faster-rcnn的话，清除Annotations里的cache里的pkl..
也评价下https://blog.csdn.net/xzzppp/article/details/52036794里解决方案
如果报错原因是数据集标注起点为(0,0)，不想改数据集的话，对Xmin,Ymin,Xmax,Ymax减一去掉可以，但是fipped操作不知道会不会还会有问题，第一条解决方案就还是别用了。。第三个解决方案，如果数据集已经改成是(1,1)为起点的话，还有这个问题就估计是部分标注Xmin和Xmax反过来了
========================
解决1问题：数据集有问题，
可能数据集没符合PASCAL VOC格式，,pascal voc的标注起始应该是1-based而不是以0开始的
首先坐标起始点都得为1，坐标不能越界
以下是检查数据集的代码：
1.检查是否图片损坏
2.检查坐标七点是否都不为0
3.检查xmax是否大于xmin，会引起
这个代码作用是检查数据集，检查是否有图片损坏（可能引起loss=NAN）,检查是否数据集开始坐标(0,0)
#改一下curDir为数据集里图片的地址和path地址
#1check if there are corrputed images in JPEG
#2check if there are '0' cordinate in xml file
#3chechk if xmin coordinates have 0 value,todo:checkif max is greater than width
#111111111111111111111111
import numpy as np
from PIL import Image
import xml.dom.minidom
import os
if __name__ == '__main__':
badFilesList = []
curDir = '/media/syy/D/repo/tf-faster-rcnn/data/VOCdevkit2007/VOC2007/JPEGImages'
for root, dirs, files in os.walk(curDir): # print(files)
# check the corrupted file in curDir
for each in files:
if each.endswith('.jpg') or each.endswith( '.JPG') or each.endswith( '.jpeg') or each.endswith( '.JPEG'): # print(each)
try:
im = Image.open(os.path.join(root, each)) # im.show()
#print(each+"is ok")
im= np.array(im,dtype=np.float32)#added recently at Feb.23
except Exception as e:
print('corrupted file:', os.path.join(root, each))
badFilesList.append(os.path.join(root, each)) # del the corrupted files
if len(badFilesList) != 0:
for each in badFilesList:
try:
#os.remove(each)
print("pretend to remove"+each)
except Exception as e:
print('Del file: %s failed, %s' % (each, e))
pass
if len(badFilesList) == 0:
print("no corrupted filesï¼already checked ",len(files))
else:
print("corrputed files are above----------------------------","already checked ",len(files))
#222=222222222222222222222222
#pause_=input()
path=r"/media/syy/D/repo/tf-faster-rcnn/data/VOCdevkit2007/VOC2007/Annotations/"
Assertion_provoker_y = []
Assertion_provoker_x = []
files = os.listdir(path)
for file in files:
file=os.path.join(path,file)
dom = xml.dom.minidom.parse(file)
root = dom.documentElement
size=dom.getElementsByTagName("size")
height=dom.getElementsByTagName("height")[0]
width=dom.getElementsByTagName("width")[0]
width=width.childNodes[0].data
height=height.childNodes[0].data
objects=dom.getElementsByTagName("object")
for object in objects:
bndbox = object.getElementsByTagName('bndbox')[0]########
xmin = bndbox.getElementsByTagName('xmin')[0]
xmin_data=xmin.childNodes[0].data
ymin = bndbox.getElementsByTagName('ymin')[0]
ymin_data=ymin.childNodes[0].data
xmax = bndbox.getElementsByTagName('xmax')[0]
xmax_data=xmax.childNodes[0].data
ymax = bndbox.getElementsByTagName('ymax')[0]
ymax_data=ymax.childNodes[0].data
#check if the xmin or yin is 0
if(int(xmin_data) ==0):
print(file,"the x of it is 0")
if(int(ymin_data) ==0):
print(file,"the y of it is 0")
##check if min is greater than max
if( int(ymin_data) > int(ymax_data) ):
Assertion_provoker_y.append(file)
if( int(xmin_data) > int(xmax_data) ):
Assertion_provoker_x.append(file)
#check if x is greater than width,or y is greater than height
if(int(xmax_data)>int(width)):
Assertion_provoker_x.append(file)
if(int(xmin_data)>int(width)):
Assertion_provoker_x.append(file)
if(int(ymax_data)>int(height)):
Assertion_provoker_y.append(file)
if(int(ymin_data)>int(height)):
Assertion_provoker_y.append(file)
print("checking assertion error done,already checked",len(files),"files,with {} y_provoker".format(len(Assertion_provoker_y)))
print("checking assertion error done,already checked",len(files),"files,with {} x_provoker".format(len(Assertion_provoker_x)))
#print(Assertion_provoker_x)
#######################################################
===========================下面是我写的处理一些越界问题的代码，比如xmin与xmax相反，最后没用上，别浪费自己的一番心血。。。贴上来（可能下面的代码会有小错误）==================================
from xml.etree.ElementTree import ElementTree,Element
import xml.dom.minidom
import os
path="/media/syy/D/repo/tf-faster-rcnn/data/VOCdevkit2007/VOC2007/Annotations/"
Assertion_provoker_x = []
Assertion_provoker_y = []
sucess_file = []
files = os.listdir(path)
for file in files:
file=os.path.join(path,file)
dom = xml.dom.minidom.parse(file)
root = dom.documentElement
objects=dom.getElementsByTagName("object")
for object in objects:
flag = False
bndbox = object.getElementsByTagName('bndbox')[0]
xmin = bndbox.getElementsByTagName('xmin')[0]
xmin_data=xmin.childNodes[0].data
xmax = bndbox.getElementsByTagName('xmax')[0]
xmax_data=xmin.childNodes[0].data
ymin = bndbox.getElementsByTagName('ymin')[0]
ymin_data=xmax.childNodes[0].data
ymax = bndbox.getElementsByTagName('ymax')[0]
ymax_data=xmax.childNodes[0].data
#----------#below is switch xmin and xmax------------------
if( int(xmin_data) > int(xmax_data) ):
Assertion_provoker_x.append(file)
xmin.childNodes[0].data = xmax_data
xmax.childNodes[0].data = xmin_data
flag = True
if( int(ymin_data) > int(ymax_data) ):
Assertion_provoker_y.append(file)
ymin.childNodes[0].data = ymax_data
ymax.childNodes[0].data = ymin_data
flag = True
#above is switch the ymin with ymax
if flag:
# with open(os.path.join(path,file),'w',encoding='UTF-8') as fh:
# dom.writexml(fh)
sucess_file.append(file)
#-------------
print("checking assertion error done,already checked",len(files),"files,with {} X_provoker".format(len(Assertion_provoker_x)),'\nand {} y_provokers '.format(len(Assertion_provoker_y)))
print("sucessfully dealt with {} files ".format(len(sucess_file)))
参考：
【1】https://blog.csdn.net/10km/article/details/64641322
【2】https://blog.csdn.net/xzzppp/article/details/52036794
=================================================================
参考：https://blog.csdn.net/hongxingabc/article/details/79048531
某些常见报错