fix: Fixed geometry utils for polygon <--> rbox conversions

[fg-mindee]

This PR introduces the following modifications:

• fixes the cv2 conventions of angle orientation for polygon <--> rbox conversions
• updates unittests

The following snippet:

import numpy as np
import matplotlib.pyplot as plt
from doctr.utils.geometry import rbbox_to_polygon

def plot_points(points):
    img = np.zeros((200, 150))
    for point in points.round().astype(int):
        img[point[1], point[0]] = 1
    plt.imshow(img); plt.show()


# Horizontal word rotated counter-clockwise by 45°
plot_points(rbbox_to_polygon((100, 100, 40, 10, 45.)))

was producing

and now produces

Additionally this snippet:

import numpy as np
import matplotlib.pyplot as plt
from doctr.utils.geometry import rbbox_to_polygon, fit_rbbox

angles = np.arange(-90, 91)
rboxes = [fit_rbbox(rbbox_to_polygon((100, 100, 40, 10, angle)))for angle in angles]
estimated_angles = [rbox[-1] for rbox in rboxes]
estimated_width = [rbox[2] for rbox in rboxes]

plt.plot(angles, estimated_angles, label="angles")
plt.plot(angles, estimated_width, label="width")
plt.xlabel("Anti-clockwise rotation")
plt.legend(loc="upper left")
plt.show()

used to yield:

and now yields (similar to before, we can't do more because of cv2):

Any feedback is welcome!

[codecov]

Codecov Report

Merging #700 (e6d9d04) into main (d67abd4) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #700      +/-   ##
==========================================
+ Coverage   96.01%   96.08%   +0.06%     
==========================================
  Files         124      124              
  Lines        4648     4648              
==========================================
+ Hits         4463     4466       +3     
+ Misses        185      182       -3     

Flag	Coverage Δ
unittests	96.08% <100.00%> (+0.06%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
doctr/utils/geometry.py	98.00% <100.00%> (ø)
doctr/models/detection/linknet/base.py	83.16% <0.00%> (-1.99%)	⬇️
...dels/detection/differentiable_binarization/base.py	92.22% <0.00%> (+2.77%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update d67abd4...e6d9d04. Read the comment docs.

[charlesmindee]

Thanks!