Create custom dataset from LabelImg

[RomRoc]

I developed a simple python script to transform annotations from LabelImg tool to Keras-Retinanet annotations format.

You can run it even in a notebook. Xml annotations image files should stay into the same DATASET_DIR.
I share it here in case it can be useful to someone else:

import xml.etree.ElementTree as ET
import numpy as np
import csv

DATASET_DIR = '/path/to/dataset'

annotations = []
classes = set([])

for xml_file in [f for f in os.listdir(DATASET_DIR) if f.endswith(".xml")]:
  tree = ET.parse(os.path.join(DATASET_DIR, xml_file))
  root = tree.getroot()

  file_name = None

  for elem in root:
    if elem.tag == 'filename':
      file_name = os.path.join(DATASET_DIR, elem.text)

    if elem.tag == 'object':
      obj_name = None
      coords = []
      for subelem in elem:
        if subelem.tag == 'name':
          obj_name = subelem.text
        if subelem.tag == 'bndbox':
          for subsubelem in subelem:
            coords.append(subsubelem.text)
      item = [file_name] + coords + [obj_name]
      annotations.append(item)
      classes.add(obj_name)

with open(ANNOTATIONS_FILE, 'w') as f:
  writer = csv.writer(f)
  writer.writerows(annotations)

with open(CLASSES_FILE, 'w') as f:
  for i, line in enumerate(classes):
    f.write('{},{}\n'.format(line,i))

[hgaiser]

Thank you for your interest and for sharing it, but I'm afraid I'll have to close this issue (as there isn't really an issue to be resolved). The issue will still be accessible, so anyone searching for this should be able to find it. 👍