converts Rotating Rectangular BBox. (OBB)

This project is aimed at converting between two representations of text detection bounding box labels (rotated rectangle bounding boxes): txt and xml.

Commonly used annotation forms for text detection include horizontal rectangular boxes, rotated rectangular boxes, arbitrary quadrilateral boxes, arbitrary polygonal boxes, and other forms (such as text snakes), etc., as shown in the figure below.

This project reads, writes and converts the rotating rectangular box representation that is most commonly used for text detection. The rotating rectangular box is different from the conventional target detection bounding box. On its basis, a rotation angle is added to rotate the conventional horizontal detection frame.

Rotated rectangular boxes usually have two representations:

1. Center point coordinates, width, height, and rotation angle are often saved in xml files.
2. The coordinates of the four corner points of the rectangle are usually simply saved directly using a txt file (such as the ICDAR2015 Dataset)

Center-rotation angle representation

Take the following detection frame as an example:

<object>
    <type>robndbox</type>
    <name>wenben</name>
    <pose>Unspecified</pose>
    <truncated>0</truncated>
    <difficult>0</difficult>
    <robndbox>
      <cx>860.5666</cx>
      <cy>734.5734</cy>
      <w>644.8657</w>
      <h>52.3775</h>
      <angle>3.031593</angle>
    </robndbox>
  </object>

Rectangular corner point coordinate representation.

Take the following detection frame as an example:

x1,y1,x2,y2,x3,y3,x4,y4,labelname

Each line in the txt file is as shown above. x1, y1, x2, y2, x3, y3, x4, y4 are the horizontal and vertical coordinates of the four corner points respectively, and labelname is the category of the detection frame.

Conversion between two representations

The specific calculation formula is as follows:

x1 = cx-w/2*cos(theta)+h/2*sin(theta)
y1 = cy-h/2*cos(theta)-w/2*sin(theta)
x2 = cx-w/2*cos(theta)+h/2*sin(theta)
y2 = cy-h/2*cos(theta)+w/2*sin(theta)
x3 = cx+w/2*cos(theta)-h/2*sin(theta)
y3 = cy+h/2*cos(theta)-w/2*sin(theta)
x4 = cx+w/2*cos(theta)+h/2*sin(theta)
y4 = cy+h/2*cos(theta)+w/2*sin(theta)

Theta is related to the angle angle. When angle is less than pi/2, theta is equal to angle. When angle is greater than pi/2, theta is equal to angle-pi.

Scrip Prepare environment

This program requires the following dependency packages:

python3.6
opencv-python 4.1.0.25
lxml 4.4.0

txt to xml

Put the image file that needs to be converted into the img folder, and the txt detection frame file into the txt folder. Run get_list.py to get the conversion list txt_to_xml_list.txt, and then run txt_to_xml.py to get the xml format tag with the same name in the xml folder document

python RoLabelImg_Transform/get_list.py --model='txt_to_xml' --input_path='./RoLabelImg_Transform/txt/'
python RoLabelImg_Transform/txt_to_xml.py

Put the image file that needs to be converted into the img folder, and the xml detection frame file into the xml folder. Run get_list.py to get the conversion list xml_to_txt_list.txt, and then run xml_to_txt.py to get the txt format tag with the same name in the txt folder. document

xml to txt

python RoLabelImg_Transform/get_list.py --model='xml_to_txt' --input_path='./RoLabelImg_Transform/xml/'
python RoLabelImg_Transform/xml_to_txt.py

txt detection frame visualization

To visualize the txt file detection frame, run the visualize.py file, and the visualized image is saved in the visualized_img folder.

python RoLabelImg_Transform/visualize.py