如何准备我的图像和注释以进行视网膜网训练？

Question

I followed this tutorial for training object detection model on coco dataset. The tutorial contains a step to download and use coco dataset and its annotations and convert them to TFRecord .

I need to use my own custom data to train, i annotated using labelimg tool which produced xml files containing (w,h,xmin,ymin,xmax,ymax) for images.

But coco dataset has JSON format with Image segmentation fields for creating a TFRecord .

Is segmentation mandatory for training resnet, retinanet?

So, can anyone guide me a procedure for creating a JSON annotation from my XML annotations without segmentation values?

xml:

<annotation>
    <folder>frames</folder>
    <filename>83.jpg</filename>
    <path>/home/tdadmin/Downloads/large/f/frames/83.jpg</path>
    <source>
        <database>Unknown</database>
    </source>
    <size>
        <width>640</width>
        <height>480</height>
        <depth>3</depth>
    </size>
    <segmented>0</segmented>
    <object>
        <name>person</name>
        <pose>Unspecified</pose>
        <truncated>0</truncated>
        <difficult>0</difficult>
        <bndbox>
            <xmin>246</xmin>
            <ymin>48</ymin>
            <xmax>350</xmax>
            <ymax>165</ymax>
        </bndbox>
    </object>
</annotation>

Answer 1

What you are doing now is kind of similar to a project I've done before. So I have some suggestions for you.

When I was training my Mask RCNN model, I used VGG Image Annotator ( you can easily find that on Google ). By using that tool, it's easy to create json annotation files. Then plug that in your training.

Hope that would help you. Feel free to comment on this if you still have questions.

Rowen

Answer 2

The annotation format actually doesn't matter. I have myself created tfrecord from txt files before. To create custom tfrecord you would have to write your own create_custom_tf_record.py just as others shown in this folder .

But since you are using coco similar annotations, you can make use of the file create_coco_tf_record.py . The important thing you need to implement yourself is the annotations_list . The annotations_list is just a dictionary so your goal is to parse your xml file into a dictionary containing key value pairs, and then pass the correct value to the feature_dict , then construct tf.train.Example from the feature_dict . Once you have the tf.train.Example created , you can create tfrecord easily.

So for your exact example, first parse the xml file.

import xml.etree.ElementTree as ET
tree = ET.parse('annotations.xml')

Then construct annotaions_list from tree like this:

annotations_list = {}
it = tree.iter()
for key in it:
    annotations_list[str(key.tag)] = key.text

Then you can create the feature_dict from the annotations_list

feature_dict = {
  'image/height':
      dataset_util.int64_feature(annotatios_list['height']),
  'image/width':
      dataset_util.int64_feature(...),
  'image/filename':
      dataset_util.bytes_feature(...),
  'image/source_id':
      dataset_util.bytes_feature(...),
  'image/key/sha256':
      dataset_util.bytes_feature(...),
  'image/encoded':
      dataset_util.bytes_feature(...),
  'image/format':
      dataset_util.bytes_feature(...),
  'image/object/bbox/xmin':
      dataset_util.float_list_feature(...),
  'image/object/bbox/xmax':
      dataset_util.float_list_feature(...),
  'image/object/bbox/ymin':
      dataset_util.float_list_feature(...),
  'image/object/bbox/ymax':
      dataset_util.float_list_feature(...),
  'image/object/class/text':
      dataset_util.bytes_list_feature(....),
  'image/object/is_crowd':
      dataset_util.int64_list_feature(...),
  'image/object/area':
      dataset_util.float_list_feature(...),
  }

Just need to make sure the feature_dict filed corresponds to the correct field from the annotations_list and the label_map .

You may wonder why exactly these fields in the feature_dict is necessary, according to the official documentation using your own dataset , the following fileds are necessary and others are optional.

'image/height': dataset_util.int64_feature(height),
  'image/width': dataset_util.int64_feature(width),
  'image/filename': dataset_util.bytes_feature(filename),
  'image/source_id': dataset_util.bytes_feature(filename),
  'image/encoded': dataset_util.bytes_feature(encoded_image_data),
  'image/format': dataset_util.bytes_feature(image_format),
  'image/object/bbox/xmin': dataset_util.float_list_feature(xmins),
  'image/object/bbox/xmax': dataset_util.float_list_feature(xmaxs),
  'image/object/bbox/ymin': dataset_util.float_list_feature(ymins),
  'image/object/bbox/ymax': dataset_util.float_list_feature(ymaxs),
  'image/object/class/text': dataset_util.bytes_list_feature(classes_text),
  'image/object/class/label': dataset_util.int64_list_feature(classes),