Question

我有一个PASCAL VOC数据集。我想用它在张量流中构建深度学习模型。我认为我需要将其转换为TFRecord文件格式以构建模型，但是我不确定我的想法是正确的。如果是的话，什么代码将PASCAL VOC转换为TFRecord文件格式。如果不是，您是否有建议加载此PASCAL VOC数据集以在Tensorflow中构建模型。这是我的PASCAL VOC数据集。

<annotation>
  <filename>000000000.jpg</filename>
  <source>
    <annotation>ArcGIS Pro 2.1</annotation>
  </source>
  <size>
    <width>256</width>
    <height>256</height>
    <depth>3</depth>
  </size>
  <object>
    <name>0</name>
    <bndbox>
        <xmin>209.62</xmin>
        <ymin>3.86</ymin>
        <xmax>256.00</xmax>
        <ymax>70.93</ymax>
    </bndbox>
 </object>
 <object>
    <name>0</name>
    <bndbox>
        <xmin>120.92</xmin>
        <ymin>126.09</ymin>
        <xmax>200.23</xmax>
        <ymax>209.97</ymax>
    </bndbox>
 </object>
 <object>
    <name>0</name>
    <bndbox>
        <xmin>237.72</xmin>
        <ymin>136.02</ymin>
        <xmax>256.00</xmax>
        <ymax>214.18</ymax>
    </bndbox>
 </object>

Answer 1

Tensorflow对象检测API为此提供了tool，您可以运行以下命令：

python object_detection/dataset_tools/create_pascal_tf_record.py \
    --label_map_path=<path/to/label/map.pbtxt> \
    --data_dir=<path/to/data/dir> --year=<year_directory_name> --set=<train|test|val> \
    --output_path=pascal_<train|test|val>.record

这需要一棵形式为

的树

data_dir
|- year_dir
   |- Annotations
      |- *.xml
   |- ImageSets
      |- Layout
        |- test.txt
        |- train.txt
        |- val.txt
        |- trainval.txt
      |- Main
        |- *.txt
   |- JPEGImages
      |- *.jpg

例如，对于通常的PASCAL数据集，结果将为：

python object_detection/dataset_tools/create_pascal_tf_record.py \
    --label_map_path=object_detection/data/pascal_label_map.pbtxt \
    --data_dir=VOCdevkit --year=VOC2012 --set=val \
    --output_path=pascal_val.record

Answer 2

VOC2007在最新的tensorflow-datasets==1.0.2版本中可用（尚未在pip上可用）。

要安装它，请在终端中运行它：

git clone https://github.com/tensorflow/datasets
cd datasets
python setup.py build
python setup.py install

用法示例（在Jupyter中的图）：

import tensorflow as tf
import tensorflow_datasets as tfds
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image, ImageDraw
%matplotlib inline

OUTLINE = (0, 255, 0)

builder = tfds.builder('voc2007')
builder.download_and_prepare()
datasets = builder.as_dataset()
train_data, test_data = datasets['train'], datasets['test']
iterator = train_data.repeat(1).batch(1).make_one_shot_iterator()
next_batch = iterator.get_next()

with tf.Session() as sess:
    for _ in range(1):
        batch = sess.run(next_batch)
        image = batch['image']
        bboxes = batch['objects']['bbox']
        bboxes, image = np.squeeze(bboxes), np.squeeze(image)
        pil_image = Image.fromarray(image.astype('uint8'), 'RGB')
        draw = ImageDraw.Draw(pil_image)
        height, width = image.shape[:2]
        try:
            if (isinstance(bboxes[0], np.float32)
                or isinstance(bboxes[0], np.float64)):
                bboxes = [bboxes]

            for bbox in bboxes:
                ymin, xmin, ymax, xmax = bbox
                xmin *= width
                xmax *= width
                ymin *= height
                ymax *= height
                c1 = (xmin, ymin)
                c2 = (xmax, ymin)
                c3 = (xmax, ymax)
                c4 = (xmin, ymax)
                draw.line([c1, c2, c3, c4, c1],
                          fill=OUTLINE,
                          width=3)
            asnumpy = np.array(pil_image)
            figure = plt.figure(figsize=tuple(x/50 for x in image.shape[:2]))
            plt.imshow(asnumpy)
        except TypeError:
            pass

如何使用xml格式的PASCAL VOC数据集在Tensorflow中构建模型

2 个答案: