目标检测：数据集的生成(PASCAL VOC Format)

一、 生成训练和测试的 filelist 以及 test_name_size

# -*- coding: utf-8 -*-

import os
import sys
import cv2
import glob
import codecs
import argparse
import numpy as np
from lxml import etree
from generate_new_xml import write_xml

reload(sys)
sys.setdefaultencoding('utf-8')


def make_voc_dir(data_save_dir, data_dir_yun):
    """
    Args:
        data_save_dir: dataset save dir, str, eg:'/home/manzp/projects/det/prep_resultdata/plate_devkit/plate'
        data_dir_yun: real dataset dir, str, eg: '/home/manzp/projects/det/datasets/plate/Image'

    Returns:
        anno_dir_f: dir of annotations
        jpg_dir_f: dir of images(soft link)
        main_txt_dir_f: annotations filename/size/obj_label/obj_bndbox txt and trainval&test img/xml pair dir
    """
    anno_dir_f = os.path.join(data_save_dir, 'Annotations')
    jpg_dir_f = os.path.join(data_save_dir, 'JPEGImages')
    main_txt_dir_f = os.path.join(data_save_dir, 'ImageSets/Main')

    if not os.path.os.path.exists(anno_dir_f):
        os.makedirs(anno_dir_f)
    if not os.path.os.path.exists(jpg_dir_f):
        os.system('ln -s {} {}'.format(data_dir_yun, jpg_dir_f))  # make soft link
    if not os.path.os.path.exists(main_txt_dir_f):
        os.makedirs(main_txt_dir_f)

    print('anno_dir is {}'.format(anno_dir_f))
    print('jpg_dir is {}'.format(jpg_dir_f))
    print('main_txt_dir is {}'.format(main_txt_dir_f))
    print('Make voc format dirs done!')
    print('####################################################')

    return anno_dir_f, jpg_dir_f, main_txt_dir_f


def extract_xml2txt(label_dir, anno_txt_dir):
    """
    Args:
        label_dir: old xml dir, eg: '/home/manzp/projects/det/datasets/plate/Label'
        anno_txt_dir: new xml info txt save dir, eg:'/home/manzp/projects/det/prep_result/data/plate_devkit/plate/ImageSets/Main'

    Returns:
        None
    """
    print('label dir is {}'.format(label_dir))
    print('anno txt save dir is {}'.format(anno_txt_dir))

    cnt = 0
    wrong_cnt = 0
    f = codecs.open(os.path.join(anno_txt_dir, 'anno.txt'), 'w', encoding='utf-8')
    for anno_xml in glob.glob(os.path.join(label_dir, '*')):
        img_path = anno_xml.replace('Label', 'Image').replace('xml', 'jpg')
        # img_path = glob.glob(anno_xml.replace('Label', 'Image').replace('xml', '*'))[0]
        img_name = img_path.split('/')[-1].encode('utf-8')
        out_info = [img_name + ' ']
        try:
            tree = etree.parse(anno_xml)  # open xml doc
            root = tree.getroot()  # get root node
            img_bgr = cv2.imdecode(np.fromfile(img_path, dtype=np.uint8), 1)
            height, width, depth = img_bgr.shape
            out_info.extend([str(height) + ' ', str(width) + ' ', str(depth) + ' '])

            for obj in root.iter('object'):
                obj_name = obj.find('name').text.encode('utf-8')
                # filter car and CAR
                if obj_name == 'car':
                    pass
                elif obj_name == 'blur':
                    plate_difficult = '1'
                    plate_obj_name = 'plate'
                    plate_coord = obj.find('polygonbox')
                    left_topx = plate_coord.find('leftTopx').text.encode('utf-8')
                    left_topy = plate_coord.find('leftTopy').text.encode('utf-8')
                    right_topx = plate_coord.find('rightTopx').text.encode('utf-8')
                    right_topy = plate_coord.find('rightTopy').text.encode('utf-8')
                    right_bottomx = plate_coord.find('rightBottomx').text.encode('utf-8')
                    right_bottomy = plate_coord.find('rightBottomy').text.encode('utf-8')
                    left_bottomx = plate_coord.find('leftBottomx').text.encode('utf-8')
                    left_bottomy = plate_coord.find('leftBottomy').text.encode('utf-8')
                    xmin_p = min(int(left_topx), int(left_bottomx))
                    xmax_p = max(int(right_topx), int(right_bottomx))
                    ymin_p = min(int(left_topy), int(right_topy))
                    ymax_p = max(int(left_bottomy), int(right_bottomy))
                    # filter invalid coordinates xml file
                    if xmin_p < 0 or xmax_p < 0 or ymin_p < 0 or ymax_p < 0:
                        break
                    elif xmin_p > int(width) or xmax_p > int(width) or ymin_p > int(height) or ymax_p > int(height):
                        break
                    elif xmin_p > xmax_p or ymin_p > ymax_p:
                        break
                    else:
                        out_info.extend([plate_obj_name + ' ', str(xmin_p) + ' ', str(ymin_p) + ' ', str(xmax_p) + ' ',
                                         str(ymax_p) + ' ',
                                         plate_difficult + ' '])
                else:
                    plate_difficult = '0'
                    plate_obj_name = 'plate'
                    plate_coord = obj.find('polygonbox')
                    left_topx = plate_coord.find('leftTopx').text.encode('utf-8')
                    left_topy = plate_coord.find('leftTopy').text.encode('utf-8')
                    right_topx = plate_coord.find('rightTopx').text.encode('utf-8')
                    right_topy = plate_coord.find('rightTopy').text.encode('utf-8')
                    right_bottomx = plate_coord.find('rightBottomx').text.encode('utf-8')
                    right_bottomy = plate_coord.find('rightBottomy').text.encode('utf-8')
                    left_bottomx = plate_coord.find('leftBottomx').text.encode('utf-8')
                    left_bottomy = plate_coord.find('leftBottomy').text.encode('utf-8')
                    xmin_p = min(int(left_topx), int(left_bottomx))
                    xmax_p = max(int(right_topx), int(right_bottomx))
                    ymin_p = min(int(left_topy), int(right_topy))
                    ymax_p = max(int(left_bottomy), int(right_bottomy))
                    # filter invalid coordinates xml file
                    if xmin_p < 0 or xmax_p < 0 or ymin_p < 0 or ymax_p < 0:
                        break
                    elif xmin_p > int(width) or xmax_p > int(width) or ymin_p > int(height) or ymax_p > int(height):
                        break
                    elif xmin_p > xmax_p or ymin_p > ymax_p:
                        break
                    else:
                        out_info.extend([plate_obj_name + ' ', str(xmin_p) + ' ', str(ymin_p) + ' ', str(xmax_p) + ' ',
                                         str(ymax_p) + ' ',
                                         plate_difficult + ' '])

            # strip the last blank and write list seq to file
            f.writelines(out_info[:-1] + [out_info[-1].strip()] + ['\n'])
            cnt += 1

        except Exception as e:
            print('Error Reason is {}'.format(e))
            print("Error: cannot parse file: {}".format(anno_xml))
            wrong_cnt += 1
            continue
    f.close()
    print('Successfully parsed {} images!'.format(cnt))
    print('Can not parse {} images!'.format(wrong_cnt))
    print('Extract xml info to anno.txt Done!')
    print('####################################################')


def write2xml(main_txt_dir, anno_dir, jpg_dir):
    """write plate obj to xml and filter no plate img
    Args:
        main_txt_dir: new xml info txt save dir
        anno_dir: new xml save dir
        jpg_dir: raw image dir

    Returns:
        None
    """
    with codecs.open(os.path.join(main_txt_dir, 'anno.txt'), 'r', encoding='utf-8') as f:
        cnt = 0
        wrong_cnt = 0
        for line in f.readlines():
            line_list = line.strip().split(' ')
            img_prefix = os.path.splitext(line_list[0])[0]
            xml_path = os.path.join(anno_dir, img_prefix + '.xml')
            img_path = os.path.join(jpg_dir, line_list[0])
            img_size = line_list[1:4]

            bndbox_list = line_list[4:]
            # filter no plate img and wrong img
            if len(bndbox_list) > 0:
                sub_bndbox_list = []
                try:
                    for i in range(0, len(bndbox_list), 6):
                        sub_bndbox = []
                        sub_bndbox.append(bndbox_list[i])  # obj_name: plate
                        sub_bndbox.append(bndbox_list[i + 1])  # xmin
                        sub_bndbox.append(bndbox_list[i + 2])  # ymin
                        sub_bndbox.append(bndbox_list[i + 3])  # xmax
                        sub_bndbox.append(bndbox_list[i + 4])  # ymax
                        sub_bndbox.append(bndbox_list[i + 5])  # plate_difficult
                        sub_bndbox_list.append(sub_bndbox)

                    write_xml(img_path, img_size, sub_bndbox_list, xml_path)
                    cnt += 1
                except Exception as e:
                    print("Error: cannot parse file: %s" % line_list[0])
                    wrong_cnt += 1
                    continue
            else:
                wrong_cnt += 1
        print('Regenerate xml wrong cnt num is {}'.format(wrong_cnt))
        print('Regenerate xml total num is {}'.format(wrong_cnt + cnt))
        print('Regenerate xml done!')
        print('####################################################')


def create_filelist(anno_dir, jpg_dir, main_txt_dir):
    """split data and generate img and xml pair list and test_name_size
    Args:
        anno_dir: new xml dir
        jpg_dir: raw image dir
        main_txt_dir: trainval test image/xml pair and test_name_size txt save dir

    Returns:
        None
    """
    xml_prefix = [i[:-4] for i in os.listdir(anno_dir)]  # strip '.xml'
    np.random.shuffle(xml_prefix)

    xml_num = len(xml_prefix)
    test_num = int(xml_num * 0.1)  # 0.1/0.9
    test_xml_prefix = xml_prefix[:test_num]
    trainval_xml_prefix = xml_prefix[test_num:]

    # generate img/xml pair list and test_name_size(prefix height width)
    train_val_path = os.path.join(main_txt_dir, 'trainval.txt')
    with codecs.open(train_val_path, 'w', encoding='utf-8') as f1:
        for prefix in trainval_xml_prefix:
            img_path = os.path.join(jpg_dir.split('data/')[1].strip(), prefix.strip() + '.jpg')
            xml_path = os.path.join(anno_dir.split('data/')[1].strip(), prefix.strip() + '.xml')
            line_pair = img_path + ' ' + xml_path + '\n'
            f1.write(line_pair)

    test_name_size_path = os.path.join(main_txt_dir, 'test_name_size.txt')
    ft = codecs.open(test_name_size_path, 'w', encoding='utf-8')
    test_path = os.path.join(main_txt_dir, 'test.txt')
    with codecs.open(test_path, 'w', encoding='utf-8') as f2:
        for prefix in test_xml_prefix:
            img_path = os.path.join(jpg_dir.split('data/')[1].strip(), prefix.strip() + '.jpg')
            xml_path = os.path.join(anno_dir.split('data/')[1].strip(), prefix.strip() + '.xml')
            line_pair = img_path + ' ' + xml_path + '\n'
            f2.write(line_pair)

            img_path_full = os.path.join(jpg_dir, prefix.strip()+'.jpg')
            img_bgr = cv2.imdecode(np.fromfile(img_path_full, dtype=np.uint8), 1)
            height, width, depth = img_bgr.shape
            ft.write(prefix + ' ' + str(height) + ' ' + str(width) + '\n')
    ft.close()

    print('Data split done!')
    print('Generate img and xml pair list done!')
    print('####################################################')


if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='manual to this script')
    parser.add_argument('--train_data_dir',
                        default='/home/manzp/projects/det/datasets/plate/Image',
                        help="the path of train_data_dir", nargs="?")
    parser.add_argument('--train_label_dir',
                        default='/home/manzp/projects/det/datasets/plate/Label',
                        help="the path of train_label_dir", nargs="?")
    parser.add_argument('--preprocess_output_dir',
                        default='/home/manzp/projects/det/plate_det_full/2.preprocess_code/prep_result1',
                        help="the path of preprocess_output_dir", nargs="?")
    args = parser.parse_args()
    DATA_DIR_YUN = args.train_data_dir
    LABEL_DIR_YUN = args.train_label_dir
    DATA_SAVE_PREFIX = 'data/plate_devkit/plate'
    DATA_SAVE_DIR = os.path.join(args.preprocess_output_dir, DATA_SAVE_PREFIX)  # the output info dir

    # 1. make voc format dir and soft link of the real data
    anno_dir, jpg_dir, main_txt_dir = make_voc_dir(data_save_dir=DATA_SAVE_DIR, data_dir_yun=DATA_DIR_YUN)

    # 2. extract filename/size/obj_label/obj_bndbox of the raw xml
    extract_xml2txt(LABEL_DIR_YUN, main_txt_dir)

    # 3. write2xml: write plate obj to xml and filter no plate img
    write2xml(main_txt_dir, anno_dir, jpg_dir)

    # 4. split data and generate img and xml pair list and test_name_size
    create_filelist(anno_dir, jpg_dir, main_txt_dir)

• generate new xml

# -*- coding: utf-8 -*-
import sys
import codecs
reload(sys)
sys.setdefaultencoding('utf-8')
from lxml import etree
from xml.etree.ElementTree import Element, SubElement
from xml.etree import ElementTree


def gen_xml(img_path, img_size):
    top = Element('annotation')
    top.set('verified', 'yes')
    folder = SubElement(top, 'folder')
    folder.text = img_path.split('/')[-3]
    filename = SubElement(top, 'filename')
    filename.text = img_path.split('/')[-1].strip()
    local_img_path = SubElement(top, 'path')
    local_img_path.text = img_path
    source = SubElement(top, 'source')
    database = SubElement(source, 'database')
    database.text = 'plate_det_full_datasets'
    size_part = SubElement(top, 'size')
    width = SubElement(size_part, 'width')
    height = SubElement(size_part, 'height')
    depth = SubElement(size_part, 'depth')
    height.text = img_size[0]
    width.text = img_size[1]
    depth.text = img_size[2]
    segmented = SubElement(top, 'segmented')
    segmented.text = '0'
    return top


def appendObjects(top, box_list):
    for each_object in box_list:
        object_item = SubElement(top, 'object')
        name = SubElement(object_item, 'name')
        name.text = each_object[0]
        flag = SubElement(object_item, 'flag')
        flag.text = 'rectangle'
        truncated = SubElement(object_item, 'truncated')
        truncated.text = '0'
        difficult = SubElement(object_item, 'difficult')
        difficult.text = each_object[5]
        bndbox = SubElement(object_item, 'bndbox')
        xmin = SubElement(bndbox, 'xmin')
        xmin.text = each_object[1]
        ymin = SubElement(bndbox, 'ymin')
        ymin.text = each_object[2]
        xmax = SubElement(bndbox, 'xmax')
        xmax.text = each_object[3]
        ymax = SubElement(bndbox, 'ymax')
        ymax.text = each_object[4]


def write_xml(img_path, img_size, box_list, xml_path):
    """
    Args:
        img_path: raw img path, str
        img_size: image size, list, eg: ['1080', '1920', '3']
        box_list: object info, list, eg: [['obj_label, 'xmin', 'ymin', 'xmax', 'ymax', 'difficult'], [...], ...]
        xml_path: new xml path, str

    Returns:
        None
    """

    root = gen_xml(img_path, img_size)
    appendObjects(root, box_list)
    rough_string = ElementTree.tostring(root, encoding='utf-8')
    root1 = etree.fromstring(rough_string)
    result = etree.tostring(root1, encoding='utf-8', pretty_print=True)
    with codecs.open(xml_path, 'w', encoding='utf-8') as out_file:
        out_file.write(result)

二、统计训练数据集各个通道上的均值（BGR 顺序）

# -*- coding: utf-8 -*-
import os
import sys
import cv2
import redis
import argparse
import numpy as np

reload(sys)
sys.setdefaultencoding('utf-8')


def compute_mean(img_dir):
    cnt = 0
    wrong_cnt = 0
    b_sum, g_sum, r_sum = 0.0, 0.0, 0.0

    for img_name in os.listdir(img_dir):
        try:
            cnt += 1
            img_path = os.path.join(img_dir, img_name)
            img_bgr = cv2.imdecode(np.fromfile(img_path, dtype=np.uint8), 1)
            h, w, c = img_bgr.shape
            b_mean1 = 1.0 * np.sum(img_bgr[:, :, 0]) / (h*w)
            g_mean1 = 1.0 * np.sum(img_bgr[:, :, 1]) / (h*w)
            r_mean1 = 1.0 * np.sum(img_bgr[:, :, 2]) / (h*w)
            b_sum += b_mean1
            g_sum += g_mean1
            r_sum += r_mean1
            print("process the: %2d example" % cnt)

        except Exception as e:
            wrong_cnt += 1
            print('wrong reason is {}'.format(e))
            print(img_name)
            continue

    b_mean = 1.0 * b_sum / cnt
    g_mean = 1.0 * g_sum / cnt
    r_mean = 1.0 * r_sum / cnt
    print('For train bgr imgs, each channel mean is {} {} {}'.format(b_mean, g_mean, r_mean))
    print('####################################################')


if __name__ == "__main__":
    # yun param
    parser = argparse.ArgumentParser(description='manual to this script')
    parser.add_argument('--train_data_dir', default='/home/manzp/datasets/det/plates/raw/plate_det_full/Image/TrainSet/Kedacom/plate_det_full',
                        help="the path of train_data_dir", nargs="?")
    parser.add_argument('--train_label_dir', default='/opt/training/DataSet/Label/TrainSet/Kedacom/dataset_name',
                        help="the path of train_label_dir", nargs="?")
    parser.add_argument('--preprocess_output_dir',
                        default='/opt/training/dl_train_plat/training_result/task-id/prep_result',
                        help="the path of preprocess_output_dir", nargs="?")
    parser.add_argument('--redis_key', default='', help='redis key', nargs="?")
    args = parser.parse_args()

    IMAGE_DIR = args.train_data_dir
    compute_mean(IMAGE_DIR)

三、通过 filelist 将 data/label 转换为 LMDB

• create_annoset.py 和 create_annoset.cpp 的用法

  # 1、create_annoset.py 的用法
  usage: create_annoset.py [-h] [--redo] [--anno-type ANNO_TYPE]
                           [--label-type LABEL_TYPE] [--backend BACKEND]
                           [--check-size] [--encode-type ENCODE_TYPE]
                           [--encoded] [--gray]
                           [--label-map-file LABEL_MAP_FILE] [--min-dim MIN_DIM]
                           [--max-dim MAX_DIM] [--resize-height RESIZE_HEIGHT]
                           [--resize-width RESIZE_WIDTH] [--shuffle]
                           [--check-label]
                           root listfile outdir exampledir
  
  # 2、create_annoset.cpp 的用法
  # ROOTFOLDER is the root folder that holds all the images and annotations,
  # and LISTFILE should be a list of files as well as their labels or label files.
  convert_annoset [FLAGS] ROOTFOLDER/ LISTFILE DB_NAME
  
  # 分类和检测 listfile 示例
  For classification task, the file should be in the format as
     imgfolder1/img1.JPEG 7
     ....
  For detection task, the file should be in the format as
     imgfolder1/img1.JPEG annofolder1/anno1.xml
     ....
  

• 转 LMDB 代码示例

  #!/bin/bash
  
  start="creat data job start!"
  end="creat data job done!"
  echo $start
  
  data_root_dir="/opt/training/dl_train_plat/training_result/convert2lmdb8/prep_result/data" # img dir, default add "/" to root directory
  txt_dir="/opt/training/dl_train_plat/training_result/convert2lmdb8/prep_result/data/plate_devkit/plate/ImageSets/Main" # txt dir
  dataset_name="plate"  # lmdb prefix
  mapfile="./labelmap_plate.prototxt"  # 要和 xml 中的类别名相对应
  anno_type="detection"
  db="lmdb"
  min_dim=0
  max_dim=0
  width=0
  height=0
  redo=1
  
  extra_cmd="--encode-type=jpg --encoded"  # The encoded image will be save in datum
  if [ $redo ]
  then
    extra_cmd="$extra_cmd --redo"
  fi
  for subset in test trainval
  do
    # last four param corresponding to create_annoset.py's root listfile outdir exampledir
    # 写上 --shuffle 和 --check-label 则为 True
    python /home/manzp/caffe_ssd/caffe/scripts/create_annoset.py \
    --anno-type=$anno_type \
    --label-map-file=$mapfile \
    --min-dim=$min_dim \
    --max-dim=$max_dim \
    --resize-width=$width \
    --resize-height=$height \
    --shuffle
    --check-label \
    $extra_cmd \
    $data_root_dir \
    $txt_dir/$subset.txt \
    $data_root_dir/$dataset_name"_"$db/$subset"_"$db \
    examples/$dataset_name
  done
  
  echo $end
  
  # labelmap_plate.prototxt
  item {
    name: "none_of_the_above"
    label: 0
    display_name: "background"
  }
  item {
    name: "plate"
    label: 1
    display_name: "plate"
  }
  

四、ReadXMLToAnnotatedDatum 源码

• gt bboxes 的左上和右下坐标除以了图像的宽和高做了归一化，并且设置了 set_difficult(difficult) 标签

// https://github.com/weiliu89/caffe/blob/ssd/src/caffe/util/io.cpp-->line 257
// Parse VOC/ILSVRC detection annotation.
bool ReadXMLToAnnotatedDatum(const string& labelfile, const int img_height,
    const int img_width, const std::map<string, int>& name_to_label,
    AnnotatedDatum* anno_datum) {
  ptree pt;
  read_xml(labelfile, pt);

  // Parse annotation.
  int width = 0, height = 0;
  try {
    height = pt.get<int>("annotation.size.height");
    width = pt.get<int>("annotation.size.width");
  } catch (const ptree_error &e) {
    LOG(WARNING) << "When parsing " << labelfile << ": " << e.what();
    height = img_height;
    width = img_width;
  }
  LOG_IF(WARNING, height != img_height) << labelfile <<
      " inconsistent image height.";
  LOG_IF(WARNING, width != img_width) << labelfile <<
      " inconsistent image width.";
  CHECK(width != 0 && height != 0) << labelfile <<
      " no valid image width/height.";
  int instance_id = 0;
  BOOST_FOREACH(ptree::value_type &v1, pt.get_child("annotation")) {
    ptree pt1 = v1.second;
    if (v1.first == "object") {
      Annotation* anno = NULL;
      bool difficult = false;
      ptree object = v1.second;
      BOOST_FOREACH(ptree::value_type &v2, object.get_child("")) {
        ptree pt2 = v2.second;
        if (v2.first == "name") {
          string name = pt2.data();
          if (name_to_label.find(name) == name_to_label.end()) {
            LOG(FATAL) << "Unknown name: " << name;
          }
          int label = name_to_label.find(name)->second;
          bool found_group = false;
          for (int g = 0; g < anno_datum->annotation_group_size(); ++g) {
            AnnotationGroup* anno_group =
                anno_datum->mutable_annotation_group(g);
            if (label == anno_group->group_label()) {
              if (anno_group->annotation_size() == 0) {
                instance_id = 0;
              } else {
                instance_id = anno_group->annotation(
                    anno_group->annotation_size() - 1).instance_id() + 1;
              }
              anno = anno_group->add_annotation();
              found_group = true;
            }
          }
          if (!found_group) {
            // If there is no such annotation_group, create a new one.
            AnnotationGroup* anno_group = anno_datum->add_annotation_group();
            anno_group->set_group_label(label);
            anno = anno_group->add_annotation();
            instance_id = 0;
          }
          anno->set_instance_id(instance_id++);
        } else if (v2.first == "difficult") {
          difficult = pt2.data() == "1";
        } else if (v2.first == "bndbox") {
          int xmin = pt2.get("xmin", 0);
          int ymin = pt2.get("ymin", 0);
          int xmax = pt2.get("xmax", 0);
          int ymax = pt2.get("ymax", 0);
          CHECK_NOTNULL(anno);
          LOG_IF(WARNING, xmin > width) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, ymin > height) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, xmax > width) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, ymax > height) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, xmin < 0) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, ymin < 0) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, xmax < 0) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, ymax < 0) << labelfile <<
              " bounding box exceeds image boundary.";
          LOG_IF(WARNING, xmin > xmax) << labelfile <<
              " bounding box irregular.";
          LOG_IF(WARNING, ymin > ymax) << labelfile <<
              " bounding box irregular.";
          // Store the normalized bounding box.
          NormalizedBBox* bbox = anno->mutable_bbox();
          bbox->set_xmin(static_cast<float>(xmin) / width);
          bbox->set_ymin(static_cast<float>(ymin) / height);
          bbox->set_xmax(static_cast<float>(xmax) / width);
          bbox->set_ymax(static_cast<float>(ymax) / height);
          bbox->set_difficult(difficult);
        }
      }
    }
  }
  return true;

五、参考资料

1、https://github.com/weiliu89/caffe/blob/ssd/scripts/create_annoset.py
2、https://github.com/weiliu89/caffe/blob/ssd/tools/convert_annoset.cpp
3、https://github.com/weiliu89/caffe/blob/ssd/src/caffe/util/io.cpp
4、一次将自己的数据集制作成PASCAL VOC格式的惨痛经历