目标检测与分类识别之提高map的方法---新生成类进行针对性训练

什么是map呢，已经有不少对map的解释说得很明白了，在这里就不再赘述。
在分类识别中，我们往往会遇到个别ap比较低的，在这种情况下，ap低的类显然会拉低map，我们在计算map的过程中，可能会得出以下结果

可以明显看出花盆、书籍纸张、金属器皿、污损用纸、垃圾桶这几个类别的ap特别低，我们可以单独对这些类别进行分析，如何将这些类别从数据集提取出来呢？可以参考这里：目标检测与分类识别之数据集分类

通过分析我们得出是数据集相对单一，图片太少，遮挡多导致的，那此时我们将这些类提取出来，生成新的数据集，方便我们对数据集进行扩充。
提取低ap的类，生成新数据集的代码如下：

import os
import xml.dom.minidom
import glob
import xml.etree.ElementTree as ET
import os,shutil

def get_categories(xml_files):
    """Generate category name to id mapping from a list of xml files.

    Arguments:
        xml_files {list} -- A list of xml file paths.

    Returns:
        dict -- category name to id mapping.
    """
    classes_names = []
    for xml_file in xml_files:
        tree = ET.parse(xml_file)
        root = tree.getroot()
        for member in root.findall("object"):
            classes_names.append(member[0].text)
    classes_names = list(set(classes_names))
    classes_names.sort()
    return {name: i for i, name in enumerate(classes_names)}

def makdir(new_datasets_path):
    if not os.path.exists(new_datasets_path):
        os.makedirs(new_datasets_path)
    print('makdir......',new_datasets_path)

def get_loss_count(countpath):
    allClass = {}
    lowApClass = {}
    with open(countpath,'r',encoding='utf-8') as f:
        lines = f.readlines()
    for i in lines:
        s=i.split(':')
        allClass[s[0]] = int(s[1])
        if int(s[1])<=400:#-------------------------------------------筛选小于400的类别
            lowApClass[s[0]] = int(s[1])
    return allClass,lowApClass






countpath = r'./VOCdevkit/count.txt'
AnnoPath = r'./VOCdevkit/VOC2007/Annotations/'
imagePath = r'./VOCdevkit/VOC2007/JPEGImages/'
newdatasets={'a':'Annotations','i':'ImageSets','j':'JPEGImages'}
newdatasetspath = r'./VOCdevkit/newdatasets_400/'
Annolist = os.listdir(AnnoPath)
makdir(newdatasetspath)
makdir(os.path.join(newdatasetspath,newdatasets['a']))
makdir(os.path.join(newdatasetspath,newdatasets['i']))
makdir(os.path.join(newdatasetspath,newdatasets['i'],'main'))
makdir(os.path.join(newdatasetspath,newdatasets['j']))
#jpglist = os.listdir(AnnoPath)
total1 = 0



total = 0
totalerror = 0
allClass = {}
lowApClass = {}
allClass,lowApClass = get_loss_count(countpath)
rate = allClass
for key ,value in rate.items():
    rate[key]=0
f=open(r'./VOCdevkit/noexit.txt','w',encoding='utf-8')
for annotation in Annolist:
    fullname = AnnoPath + annotation
    print(fullname,annotation)

    dom = xml.dom.minidom.parse(fullname) # 打开XML文件
    # print(fullname,annotation)
    # print('---------------', total1)

    collection = dom.documentElement # 获取元素对象
    objectlist = collection.getElementsByTagName('object') # 获取标签名为object的信息
    filename = collection.getElementsByTagName('filename')  # 获取标签名为object的信息
    # name = filename[:-4]
    for object in objectlist:
        namelist = object.getElementsByTagName('name') # 获取子标签name的信息
        objectname = namelist[0].childNodes[0].data # 取到name具体的值
        # rate[objectname] += 1
        if objectname in lowApClass.keys(): # 判断字典里有没有标签，如无添加相应字段
            if   os.path.exists(os.path.join(imagePath,filename[0].childNodes[0].data)):
                shutil.copy(os.path.join(AnnoPath, annotation), os.path.join(newdatasetspath, newdatasets['a']))
                shutil.copy(os.path.join(imagePath, filename[0].childNodes[0].data),
                            os.path.join(newdatasetspath, newdatasets['j']))
                total1 += 1
            else:
                totalerror+=1
                print( '-------------------路径不存在个数', totalerror,annotation)
                f.write( os.path.join(AnnoPath, annotation)+'\n')

        rate[objectname] += 1
    total += 1
f.close()
print(rate)
print(total1,total,'-------------------路径不存在个数',totalerror)

将代码保存为脚本get_low_ap_class.py
运行可得到新的数据集

然后我们就可以针对这些数据集进行扩充或者做其他处理，来提升其ap

如果新生成的类遇到标签与图片不对应的情况，可以执行以下脚本

import os
import shutil


img_list = os.listdir('./VOCdevkit/VOC2007/JPEGImages')

anno_list = os.listdir('./VOCdevkit/VOC2007/Annotations')

#anno_name = []
#for anno in anno_list:
#	anno_name.append(anno.split('.')[0])
#for i in range(len(img_list)):
#	img_name = img_list[i].split('.')[0]
#	if img_name not in anno_name:
#		print(img_name)
                                          # --------------------------->
#		shutil.copy('./Annotations/' + img_name + '.xml', './VOCdevkit/VOC2007/Annotations/' + img_name + '.xml')

img_name = []
for img in img_list:
	img_name.append(img.split('.')[0])

for i in range(len(anno_list)):
	anno_name = anno_list[i].split('.')[0]
	if anno_name not in img_name:
		print(anno_name)

		shutil.copy('K:/VOCdevkit/VOC2007/JPEGImages/' + anno_name + '.jpg', './VOCdevkit/VOC2007/JPEGImages/' + anno_name + '.jpg')