将 VOC 格式转换成 COCO 格式

  2. import os.path as osp
  3. import xml.etree.ElementTree as ET
  5. import mmcv
  7. def voc_classes():
  8.     return [
  9.         'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat',
  10.         'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person',
  11.         'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
  12.     ]    
  14. from glob import glob
  15. from tqdm import tqdm
  16. from PIL import Image
  17. label_ids = {name: i + 1 for i, name in enumerate(voc_classes())}
  20. def get_segmentation(points):
  22.     return [points[0], points[1], points[2] + points[0], points[1],
  23.              points[2] + points[0], points[3] + points[1], points[0], points[3] + points[1]]
  26. def parse_xml(xml_path, img_id, anno_id):
  27.     tree = ET.parse(xml_path)
  28.     root = tree.getroot()
  29.     annotation = []
  30.     for obj in root.findall('object'):
  31.         name = obj.find('name').text
  32.         if name == 'waterweeds':
  33.             continue
  34.         category_id = label_ids[name]
  35.         bnd_box = obj.find('bndbox')
  36.         xmin = int(bnd_box.find('xmin').text)
  37.         ymin = int(bnd_box.find('ymin').text)
  38.         xmax = int(bnd_box.find('xmax').text)
  39.         ymax = int(bnd_box.find('ymax').text)
  40.         w = xmax - xmin + 1
  41.         h = ymax - ymin + 1
  42.         area = w*h
  43.         segmentation = get_segmentation([xmin, ymin, w, h])
  44.         annotation.append({
  45.                         "segmentation": segmentation,
  46.                         "area": area,
  47.                         "iscrowd": 0,
  48.                         "image_id": img_id,
  49.                         "bbox": [xmin, ymin, w, h],
  50.                         "category_id": category_id,
  51.                         "id": anno_id,
  52.                         "ignore": 0})
  53.         anno_id += 1
  54.     return annotation, anno_id
  57. def cvt_annotations(img_path, xml_path, out_file):
  58.     images = []
  59.     annotations = []
  61.     # xml_paths = glob(xml_path + '/*.xml')
  62.     img_id = 1
  63.     anno_id = 1
  64.     for img_path in tqdm(glob(img_path + '/*.jpg')):
  65.         w, h = Image.open(img_path).size
  66.         img_name = osp.basename(img_path)
  67.         img = {"file_name": img_name, "height": int(h), "width": int(w), "id": img_id}
  68.         images.append(img)
  70.         xml_file_name = img_name.split('.')[0] + '.xml'
  71.         xml_file_path = osp.join(xml_path, xml_file_name)
  72.         annos, anno_id = parse_xml(xml_file_path, img_id, anno_id)
  73.         annotations.extend(annos)
  74.         img_id += 1
  76.     categories = []
  77.     for k,v in label_ids.items():
  78.         categories.append({"name": k, "id": v})
  79.     final_result = {"images": images, "annotations": annotations, "categories": categories}
  80.     mmcv.dump(final_result, out_file)
  81.     return annotations
  84. def main():
  85.     xml_path = r'E:\VOC\VOC2007/Annotations'  # xml 所在文件夹
  86.     img_path = r'E:\VOC\VOC2007/JPEGImages' # 图片所在文件夹
  87.     print('processing {} ...'.format("xml format annotations"))
  88.     cvt_annotations(img_path, xml_path, r'E:/VOC/all.json') # 保存的结果
  89.     print('Done!')
  92. if __name__ == '__main__':
  93.     main()

数据分析

VOC2007 总共有 9963 张图片，30638 个标注框
官方的 VOC2007 将其进行划分：训练集（2501张图片）、验证集（2510张图片）、测试集（4952张图片）

每个类别的标注框数量如下所示，可以看到训练集和测试集上每个类别的标注框数量是相等的（保证训练集和验证集的分布是一致的）

image.png
整个数据集（9963 张图片）的类别分布如下，可以看出，person 这个类别是最多的
image.png

data_analysis-Copy1.ipynb