feat: Add YOLO_OBB

[ftapajos]

Could not test this commit. Any help is appreciated

ref: https://docs.ultralytics.com/datasets/obb/#supported-obb-dataset-formats ref: #280

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 90.69767% with 4 lines in your changes are missing coverage. Please review.

:exclamation: No coverage uploaded for pull request base (master@5171df4). Click here to learn what that means.

Files	Patch %	Lines
label_studio_converter/converter.py	89.47%	2 Missing :warning:
label_studio_converter/main.py	0.00%	2 Missing :warning:

Additional details and impacted files

@@            Coverage Diff            @@
##             master     #281   +/-   ##
=========================================
  Coverage          ?   49.11%           
=========================================
  Files             ?       22           
  Lines             ?     1859           
  Branches          ?        0           
=========================================
  Hits              ?      913           
  Misses            ?      946           
  Partials          ?        0           

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[makseq]

Could you please add tests?

[makseq]

Now it looks almost awesome! :-) The last thing I would add - a few lines in the test where we test for really converted coordinates in the YOLO output, is it possible?

[ftapajos]

@makseq I can do it, but I do need some help in order to generate a practical example of annotated data in YOLO_OBB format, so that I can be sure it is working. Documentation in https://docs.ultralytics.com/datasets/obb/#supported-obb-dataset-formats is pretty misty. I tried executing

 python label_studio_converter/main.py export -i <exported_task>.json -o output_dir --image-dir <image_dir>  -f YOLO

but I get the error:

label_studio_converter/converter.py", line 779, in convert_to_yolo
    data_key = self._data_keys[0]

Since not even YOLO is working in master branch, I assume that is not a problem with my changes. Also, when I determine -c file, I don't get this error, but I'm not sure what the .xml file should be or how could I export it from label-studio

[makseq]

@ftapajos You need to add labeling config:

python label_studio_converter/main.py export -c path/to/label_config.xml -i <exported_task>.json -o output_dir --image-dir <image_dir>  -f YOLO

[makseq]

---

I do need some help in order to generate a practical example of annotated data in YOLO_OBB format

Let's try to cover it at least with a simple approach like this - just manually calculate one bbox:

Label studio bbox:

{
            "x": 7.430746538901997,
            "y": 80.15868310961756,
            "width": 39.142525358447735,
            "height": 44.06258937962363,
            "rotation": 294.75527908528767
}

Yolo obb:

1 0.07430746538901997 0.8015868310961756 0.23821421072904478 1.1570419027658772 0.6383486098885134 0.9725327135827279 0.4744418645484886 0.6170776419130263

=> top left = 0.07430746538901997 0.8015868310961756 top right = 0.23821421072904478 1.1570419027658772 bottom left = 0.6383486098885134 0.9725327135827279 bottom right = 0.4744418645484886 0.6170776419130263

and it seems these coordinates are wrong..

[makseq]

Try using this Label config =

<View>
  <Image name="image" value="$image"/>
  <RectangleLabels name="label" toName="image">
    <Label value="Airplane" background="green"/>
    <Label value="Car" background="blue"/>
  </RectangleLabels>
</View>

Task =

{
  "id": 1,
  "data": {
    "image": "https://data.heartex.net/open-images/train_0/mini/0096ba8fb44c1e2a.jpg"
  },
  "annotations": [
    {
      "result": [
        {
          "original_width": 768,
          "original_height": 578,
          "image_rotation": 0,
          "value": {
            "x": 1.2027596671261573,
            "y": 79.84292731596452,
            "width": 31.60572787517726,
            "height": 22.9604606991428,
            "rotation": 319.9085956625488,
            "rectanglelabels": [
              "Car"
            ]
          },
          "id": "oid3",
          "from_name": "label",
          "to_name": "image",
          "type": "rectanglelabels",
        }
      ]
    }
  ]
}

and check this piece of code, it works correctly and visualizes the rectangle from the task above:

# pip install opencv-python numpy matplotlib
import cv2
import math
import numpy as np
import matplotlib.pyplot as plt


def get_rotated_corner_points(box):
    """Get the corner points of a rotated bounding box.
    Note: It's important to take into account
    the original width and height of the image to get the correct coordinates.

    Args:
        box (dict): The bounding box with the following keys:
            - x: The x-coordinate of the top left corner (in percentage, 0-100)
            - y: The y-coordinate of the top left corner (in percentage, 0-100)
            - width: The width of the box (in percentage, 0-100)
            - height: The height of the box (in percentage, 0-100)
            - rotation: The rotation angle in degrees (optional)
            - original_width: The original width of the image (in pixels)
            - original_height: The original height of the image  (in pixels)
    """
    image_width, image_height = box["original_width"], box["original_height"]
    w, h = box["width"] * image_width / 100, box["height"] * image_height / 100
    a = math.pi * (box["rotation"] / 180.0) if "rotation" in box else 0.0
    cos_a, sin_a = math.cos(a), math.sin(a)

    x1, y1 = box["x"] * image_width / 100, box["y"] * image_height / 100  # top left
    x2, y2 = x1 + w * cos_a, y1 + w * sin_a  # top right
    x3, y3 = x2 - h * sin_a, y2 + h * cos_a  # bottom right
    x4, y4 = x1 - h * sin_a, y1 + h * cos_a  # bottom left

    return [(x1, y1), (x2, y2), (x3, y3), (x4, y4)]


# Load an example image for background (this is just for visualization)
# download this image here: https://data.heartex.net/open-images/train_0/mini/0096ba8fb44c1e2a.jpg
image = cv2.imread("0096ba8fb44c1e2a.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = image

label = {
    "x": 1.2027596671261573,
    "y": 79.84292731596452,
    "width": 31.60572787517726,
    "height": 22.9604606991428,
    "rotation": 319.9085956625488,
    "original_width": 768,
    "original_height": 578,
}
points = get_rotated_corner_points(label)

# test output points
assert points == [
    (9.237194243528888, 461.4921198862749),
    (194.9315420018524, 305.17056596785676),
    (280.3989008248278, 406.69722722112033),
    (94.70455306650432, 563.0187811395385),
]

for start, end in zip(points, points[1:] + points[:1]):
    cv2.line(image, tuple(np.int0(start)), tuple(np.int0(end)), (255, 0, 0), 2)


# Display the image
plt.figure(figsize=(10, 10))
plt.imshow(image)
plt.axis("off")
plt.show()

It would be great if you could reuse get_rotated_corner_points(). Also you can reuse assert for the test.

[ftapajos]

Thanks @makseq, after the hints you gave me, I could manage to correct the commit changing almost everything in order to get it to work with my dataset . Should I assert the same results you provided?

[makseq]

@ftapajos yes, use it in assert. Also I've put some comment in your code, it's not very well formatted.

[makseq]

@ftapajos please let me know, would you like to finish your contribution?

[ftapajos]

@makseq I would, but to be fair I'm a little out of time until the end of the next week. I really don't mind if anyone wants to takeover this PR until then

[makseq]

Closed because of this similar PR: https://github.com/HumanSignal/label-studio-converter/pull/287