Reading PASCAL VOC annotations in python

Question

I have annotations in xml files such as this one, which follows the PASCAL VOC convention:

<annotation>
<folder>training</folder>
<filename>chanel1.jpg</filename>
<source>
<database>synthetic initialization</database>
<annotation>PASCAL VOC2007</annotation>
<image>synthetic</image>
<flickrid>none</flickrid>
</source>
<owner>
<flickrid>none</flickrid>
<name>none</name>
</owner>
<size>
<width>640</width>
<height>427</height>
<depth>3</depth>
</size>
<segmented>0</segmented>
<object>
<name>chanel</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>344</xmin>
<ymin>10</ymin>
<xmax>422</xmax>
<ymax>83</ymax>
</bndbox>
</object>
<object>
<name>chanel</name>
<pose>Unspecified</pose>
<truncated>0</truncated>
<difficult>0</difficult>
<bndbox>
<xmin>355</xmin>
<ymin>165</ymin>
<xmax>443</xmax>
<ymax>206</ymax>
</bndbox>
</object>
</annotation>

What is the cleanest way of retrieving for example the fields filename and bndbox in Python?

I was trying to ElementTree, which seems to be the official Python solution, but I can't make it work.

My code so far:

from xml.etree import ElementTree as ET
tree = ET.parse("data/all/annotations/" + file)
fn = tree.find('filename').text
boxes = tree.findall('bndbox')

this produces

fn == 'chanel1.jpg'
boxes == []

So it succesfully extracts the filename field, but not the bndbox'es.

Solution

That's a quite easy solution for your problem:

This will return your box coordinates in a nested list [xmin, ymin, xmax, ymax] and the filename Once I struggled with bndbox tags which where mixed up (ymin, xmin,...) or any other strange combinations, so this code read the tags not only the position.

Finally I updated the code. Thanks to craq and Pritesh Gohil, you were absolutely right.

Hope it helps...

import xml.etree.ElementTree as ET


def read_content(xml_file: str):

    tree = ET.parse(xml_file)
    root = tree.getroot()

    list_with_all_boxes = []

    for boxes in root.iter('object'):

        filename = root.find('filename').text

        ymin, xmin, ymax, xmax = None, None, None, None

        ymin = int(boxes.find("bndbox/ymin").text)
        xmin = int(boxes.find("bndbox/xmin").text)
        ymax = int(boxes.find("bndbox/ymax").text)
        xmax = int(boxes.find("bndbox/xmax").text)

        list_with_single_boxes = [xmin, ymin, xmax, ymax]
        list_with_all_boxes.append(list_with_single_boxes)

    return filename, list_with_all_boxes

name, boxes = read_content("file.xml")

Comments

[Waylon Flinn]

Great answer. Very useful. One small correction, bounding box coordinates should be offset by -1 (that is, you need to subtract 1 from coordinates when assigning to variables). References: gluon-cv.mxnet.io/_modules/gluoncv/data/pascal_voc/…

[craq]

how come you use boxes.findall? I would expect just one bounding box per object.

[Pritesh Gohil]

Exactly @craq, Second for loop is not required at all. Remove'for box in boxes.findall("bndbox"): and modify ymin = int(boxes.find("bndbox/ymin").text) xmin = int(boxes.find("bndbox/xmin").text) ymax = int(boxex.find("bndbox/ymax").text) xmax = int(boxex.find("bndbox/xmax").text)

[ebk]

@WaylonFlinn Could you please elaborate on why are the -1 offsets needed? The code in the link has no comment for the part.