摘要

这篇文章主要讲如何从VOC和COCO数据集中提取特定的类，比如人。我们想做个行人检测的项目，需要从一些公开的数据集中提取一些行人的数据做补充。

1、提取VOC数据集

# -*- coding: utf-8 -*-
# @Function:There are 20 classes in VOC data set. If you need to extract specific classes, you can use this program to extract them.

import os
import shutil

ann_filepath = r"./VOCdevkit/VOC2012/Annotations/"
img_filepath = r"./VOCdevkit/VOC2012/JPEGImages/"
img_savepath = r"./VOCdevkit/VOC2012/tte/"
ann_savepath = r"./VOCdevkit/VOC2012/xml/"

if not os.path.exists(img_savepath):
    os.mkdir(img_savepath)

if not os.path.exists(ann_savepath):
    os.mkdir(ann_savepath)
names = locals()
classes = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle',
           'bus', 'car', 'cat', 'chair', 'cow', 'diningtable',
           'dog', 'horse', 'motorbike', 'pottedplant',
           'sheep', 'sofa', 'train', 'tvmonitor', 'person']

for file in os.listdir(ann_filepath):
    print(file)
    fp = open(ann_filepath + '\\' + file)
    ann_savefile = ann_savepath + file
    fp_w = open(ann_savefile, 'w')
    lines = fp.readlines()

    ind_start = []
    ind_end = []
    lines_id_start = lines[:]
    lines_id_end = lines[:]

    classes5 = '\t\t<name>person</name>\n'

    # 在xml中找到object块，并将其记录下来
    while "\t<object>\n" in lines_id_start:
        a = lines_id_start.index("\t<object>\n")
        ind_start.append(a)
        lines_id_start[a] = "delete"

    while "\t</object>\n" in lines_id_end:
        b = lines_id_end.index("\t</object>\n")
        ind_end.append(b)
        lines_id_end[b] = "delete"

    # names中存放所有的object块
    i = 0
    for k in range(0, len(ind_start)):
        names['block%d' % k] = []
        for j in range(0, len(classes)):
            if classes[j] in lines[ind_start[i] + 1]:
                a = ind_start[i]
                for o in range(ind_end[i] - ind_start[i] + 1):
                    names['block%d' % k].append(lines[a + o])
                break
        i += 1
        # print(names['block%d' % k])

    # xml头
    string_start = lines[0:ind_start[0]]
    # xml尾
    string_end = [lines[len(lines) - 1]]

    # 在给定的类中搜索，若存在则，写入object块信息
    a = 0
    for k in range(0, len(ind_start)):

        if classes5 in names['block%d' % k]:
            a += 1
            string_start += names['block%d' % k]
    string_start += string_end
    for c in range(0, len(string_start)):
        fp_w.write(string_start[c])
    fp_w.close()
    # 如果没有我们寻找的模块，则删除此xml，有的话拷贝图片
    if a == 0:
        os.remove(ann_savepath + file)
    else:
        name_img = img_filepath + os.path.splitext(file)[0] + ".jpg"
        shutil.copy(name_img, img_savepath)
    fp.close()

2、从COCO中提取特定的类别

从coco数据集中提取特定的类，并将其转为VOC格式的xml文件保存。

from pycocotools.coco import COCO
import os
import shutil
from tqdm import tqdm
import skimage.io as io
import matplotlib.pyplot as plt
import cv2
from PIL import Image, ImageDraw

# the path you want to save your results for coco to voc
savepath = "./coco2017/result/"
img_dir = savepath + 'images/'
anno_dir = savepath + 'Annotations/'
datasets_list = ['train2017','val2017']

classes_names = ['person']
dataDir = './coco2017/annotations_trainval2017'

headstr = """\
<annotation>
    <folder>VOC</folder>
    <filename>%s</filename>
    <source>
        <database>My Database</database>
        <annotation>COCO</annotation>
        <image>flickr</image>
        <flickrid>NULL</flickrid>
    </source>
    <owner>
        <flickrid>NULL</flickrid>
        <name>company</name>
    </owner>
    <size>
        <width>%d</width>
        <height>%d</height>
        <depth>%d</depth>
    </size>
    <segmented>0</segmented>
"""
objstr = """\
    <object>
        <name>%s</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>%d</xmin>
            <ymin>%d</ymin>
            <xmax>%d</xmax>
            <ymax>%d</ymax>
        </bndbox>
    </object>
"""

tailstr = '''\
</annotation>
'''

# if the dir is not exists,make it,else delete it
def mkr(path):
    if os.path.exists(path):
        shutil.rmtree(path)
        os.makedirs(path)
    else:
        os.makedirs(path)

mkr(img_dir)
mkr(anno_dir)

def id2name(coco):
    classes = dict()
    for cls in coco.dataset['categories']:
        classes[cls['id']] = cls['name']
    return classes

def write_xml(anno_path, head, objs, tail):
    f = open(anno_path, "w")
    f.write(head)
    for obj in objs:
        f.write(objstr % (obj[0], obj[1], obj[2], obj[3], obj[4]))
    f.write(tail)

def save_annotations_and_imgs(coco, dataset, filename, objs):
    # eg:COCO_train2014_000000196610.jpg-->COCO_train2014_000000196610.xml
    anno_path = anno_dir + filename[:-3] + 'xml'
    img_path = dataDir + '/'+dataset + '/' + filename
    print("img_path:",img_path)
    dst_imgpath = img_dir + filename

    img = cv2.imread(img_path)
    if (img.shape[2] == 1):
        print(filename + " not a RGB image")
        return
    shutil.copy(img_path, dst_imgpath)

    head = headstr % (filename, img.shape[1], img.shape[0], img.shape[2])
    tail = tailstr
    write_xml(anno_path, head, objs, tail)

def showimg(coco, dataset, img, classes, cls_id, show=True):
    global dataDir
    I = Image.open('%s/%s/%s' % (dataDir, dataset, img['file_name']))
    # 通过id，得到注释的信息
    annIds = coco.getAnnIds(imgIds=img['id'], catIds=cls_id, iscrowd=None)
    # print(annIds)
    anns = coco.loadAnns(annIds)
    # print(anns)
    # coco.showAnns(anns)
    objs = []
    for ann in anns:
        class_name = classes[ann['category_id']]
        if class_name in classes_names:
            print(class_name)
            if 'bbox' in ann:
                bbox = ann['bbox']
                xmin = int(bbox[0])
                ymin = int(bbox[1])
                xmax = int(bbox[2] + bbox[0])
                ymax = int(bbox[3] + bbox[1])
                obj = [class_name, xmin, ymin, xmax, ymax]
                objs.append(obj)
                draw = ImageDraw.Draw(I)
                draw.rectangle([xmin, ymin, xmax, ymax])
    if show:
        plt.figure()
        plt.axis('off')
        plt.imshow(I)
        plt.show()

    return objs

for dataset in datasets_list:
    # ./COCO/annotations/instances_train2014.json
    annFile = '{}/annotations/instances_{}.json'.format(dataDir, dataset)

    # COCO API for initializing annotated data
    coco = COCO(annFile)
    '''
    COCO 对象创建完毕后会输出如下信息:
    loading annotations into memory...
    Done (t=0.81s)
    creating index...
    index created!
    至此, json 脚本解析完毕, 并且将图片和对应的标注数据关联起来.
    '''
    # show all classes in coco
    classes = id2name(coco)
    print(classes)
    # [1, 2, 3, 4, 6, 8]
    classes_ids = coco.getCatIds(catNms=classes_names)
    print(classes_ids)
    for cls in classes_names:
        # Get ID number of this class
        cls_id = coco.getCatIds(catNms=[cls])
        img_ids = coco.getImgIds(catIds=cls_id)
        print(cls, len(img_ids))
        # imgIds=img_ids[0:10]
        for imgId in tqdm(img_ids):
            img = coco.loadImgs(imgId)[0]
            filename = img['file_name']
            print("filename:",filename)
            print("dataset:",dataset)
            objs = showimg(coco, dataset, img, classes, classes_ids, show=False)
            print(objs)
            save_annotations_and_imgs(coco, dataset, filename, objs)

转载自：https://wanghao.blog.csdn.net/article/details/127387556