tensorflow对象检测API：生成自定义数据集的TF记录

Question

I am trying to retrain the tensorflow object detection API with my own data i have labelled my image with labelImg but when i am using the script create_pascal_tf_record.py which is included in the tensorflow/models/research, i got some errors and i dont really know why it happens

python object_detection/dataset_tools/create_pascal_tf_record.py --data_dir=/home/jim/Documents/tfAPI/workspace/training_cabbage/images/train/ --label_map_path=/home/jim/Documents/tfAPI/workspace/training_cabbage/annotations/label_map.pbtxt --output_path=/home/jim/Desktop/cabbage_pascal.record --set=train --annotations_dir=/home/jim/Documents/tfAPI/workspace/training_cabbage/images/train/ --year=merged
Traceback (most recent call last):
  File "object_detection/dataset_tools/create_pascal_tf_record.py", line 185, in <module>
    tf.app.run()
  File "/home/jim/.virtualenvs/enrouteDeepDroneTF/local/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 125, in run
    _sys.exit(main(argv))
  File "object_detection/dataset_tools/create_pascal_tf_record.py", line 167, in main
    examples_list = dataset_util.read_examples_list(examples_path)
  File "/home/jim/Documents/tfAPI/models/research/object_detection/utils/dataset_util.py", line 59, in read_examples_list
    lines = fid.readlines()
  File "/home/jim/.virtualenvs/enrouteDeepDroneTF/local/lib/python2.7/site-packages/tensorflow/python/lib/io/file_io.py", line 188, in readlines
    self._preread_check()
  File "/home/jim/.virtualenvs/enrouteDeepDroneTF/local/lib/python2.7/site-packages/tensorflow/python/lib/io/file_io.py", line 85, in _preread_check
    compat.as_bytes(self.__name), 1024 * 512, status)
  File "/home/jim/.virtualenvs/enrouteDeepDroneTF/local/lib/python2.7/site-packages/tensorflow/python/framework/errors_impl.py", line 528, in __exit__
    c_api.TF_GetCode(self.status.status))
tensorflow.python.framework.errors_impl.NotFoundError: /home/jim/Documents/tfAPI/workspace/training_cabbage/images/train/VOC2007/ImageSets/Main/aeroplane_train.txt; No such file or directory

the train folder contains the xml and the jpg
the annotation folder contains my labelmap.pbtxt for my custom class
and i want to publish the TF record file on the desktop

it seems that it cant find a file in my images and annotations folder but i dont know why If someone has idea, thank you in advance

Answer 1

This error happens because you use the code for PASCAL VOC, which requires certain data folders structure. Basically, you need to download and unpack VOCdevkit to make the script work. As user phd pointed you out, you need the file VOC2007/ImageSets/Main/aeroplane_train.txt .

I recommend you to write your own script for tfrecords creation, it's not difficult. You need just two key components:

• Loop over your data that reads the images and annotations
• A function that encodes the data into tf.train.Example . For that you can pretty much re-use the dict_to_tf_example

Inside the loop, having created the tf_example , pass it to TFRecordWriter :

writer.write(tf_example.SerializeToString())

Answer 2

OK for future references, this is how i add background images to the dataset allowing the model to train on it. Functions used from: datitran/raccoon_dataset

1. Generate CSV file -> xml_to_csv.py
2. Generate TFRecord from CSV file -> generate_tfrecord.py

First Step - Creating XML file for it

Example of background image XML file

<annotation>
    <folder>test/</folder>
    <filename>XXXXXX.png</filename>
    <path>your_path/test/XXXXXX.png</path>
    <source>
        <database>Unknown</database>
    </source>
    <size>
        <width>640</width>
        <height>640</height>
        <depth>3</depth>
    </size>
    <segmented>0</segmented>
</annotation>

Basically you remove the entire <object> (ie no annotation )

Second Step - Generate CSV file

Using the xml_to_csv.py I just add a little change, to consider the XML file that do not have any annotation (the background images) as so: From the original: https://github.com/datitran/raccoon_dataset/blob/93938849301895fb73909842ba04af9b602f677a/xml_to_csv.py#L12-L22

I add:

value = None
for member in root.findall('object'):
            value = (root.find('filename').text,
                     int(root.find('size')[0].text),
                     int(root.find('size')[1].text),
                     member[0].text,
                     int(member[4][0].text),
                     int(member[4][1].text),
                     int(member[4][2].text),
                     int(member[4][3].text)
                     )
            xml_list.append(value)
        if value is None:
            value = (root.find('filename').text,
                     int(root.find('size')[0].text),
                     int(root.find('size')[1].text),
                     '-1',
                     '-1',
                     '-1',
                     '-1',
                     '-1'
                     )
            xml_list.append(value)

I'm just adding negative values to the coordinates of the bounding box if there is no in the XML file, which is the case for the background images, and it will be usefull when generating the TFRecords.

Third and Final Step - Generating the TFRecords

Now, when creating the TFRecords, if the the corresponding row/image has negative coordinates, i just add zero values to the record (before, this would not even be possible).

So from the original: https://github.com/datitran/raccoon_dataset/blob/93938849301895fb73909842ba04af9b602f677a/generate_tfrecord.py#L60-L66

I add:

for index, row in group.object.iterrows():
        if int(row['xmin']) > -1:
            xmins.append(row['xmin'] / width)
            xmaxs.append(row['xmax'] / width)
            ymins.append(row['ymin'] / height)
            ymaxs.append(row['ymax'] / height)
            classes_text.append(row['class'].encode('utf8'))
            classes.append(class_text_to_int(row['class']))
        else:
            xmins.append(0)
            xmaxs.append(0)
            ymins.append(0)
            ymaxs.append(0)
            classes_text.append('something'.encode('utf8'))  # this doe not matter for the background
            classes.append(5000)

To note that in the class_text (of the else statement), since for the background images there are no bounding boxes, you can replace the string with whatever you would like, for the background cases, this will not appear anywhere.

And lastly for the classes (of the else statement) you just need to add a number label that does not belong to neither of your own classes.

For those who are wondering, I've used this procedure many times, and currently works for my use cases.

Hope it helped in some way.