refactor: Cleaned rotation transforms

[fg-mindee]

This PR introduces the following modifications:

• updates interface of RandomRotate to directly take boxes as target and not a dict
• cleans the codebase to properly handle the expand option
• updates unittests
• added illustration in the documentation

Please note that using expand=True changes the size of the output image.

Now, the following snippet:

import cv2
import numpy as np
import matplotlib.pyplot as plt
from doctr.file_utils import is_tf_available
from doctr.datasets import FUNSD
from doctr.transforms.functional import rotate

# Get some data
ds = FUNSD()
img, target = ds[0]
# Perform rotation
rot_img, rot_boxes = rotate(img, target['boxes'], 30)
rotbis_img, rotbis_boxes = rotate(img, target['boxes'], 30, expand=True)

# Create binary map for visualization
init_target = np.zeros(img.shape[:2] if is_tf_available() else img.shape[-2:], dtype=bool)
for box in target['boxes']:
    init_target[box[1]: box[3], box[0]: box[2]] = True

new_target = np.zeros(rot_img.shape[:2] if is_tf_available() else rot_img.shape[-2:], dtype=np.uint8)

for box in rot_boxes:
    box = cv2.boxPoints(((int(box[0]), int(box[1])), (int(box[2]), int(box[3])), -box[4]))
    box = np.int0(box)
    cv2.fillPoly(new_target, [box], 1)

exp_target = np.zeros(rotbis_img.shape[:2] if is_tf_available() else rotbis_img.shape[-2:], dtype=np.uint8)

for box in rotbis_boxes:
    box = cv2.boxPoints(((int(box[0]), int(box[1])), (int(box[2]), int(box[3])), -box[4]))
    box = np.int0(box)
    cv2.fillPoly(exp_target, [box], 1)

# Make it work for PyTorch as well
if not is_tf_available():
    img = img.permute(1, 2, 0)
    rot_img = rot_img.permute(1, 2, 0)
    rotbis_img = rotbis_img.permute(1, 2, 0)

# Plot everything
_, axes = plt.subplots(2, 3)
axes[0, 0].imshow(img)
axes[0, 0].set_title('Original image')
axes[0, 1].imshow(rot_img)
axes[0, 1].set_title('Rotated image (expand=False)')
axes[0, 2].imshow(rotbis_img)
axes[0, 2].set_title('Rotated image (expand=True)')
axes[1, 0].imshow(init_target)
axes[1, 1].imshow(new_target)
axes[1, 2].imshow(exp_target)

for axs in axes:
    for ax in axs:
        ax.axis('off')

plt.show()

yields:

for both Pytorch & TF

The documentation of RandomRotate will now render as follows:

Any feedback is welcome!

[charlesmindee]

thanks !

[codecov]

Codecov Report

Merging #536 (13e6b3b) into main (323d484) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##             main     #536   +/-   ##
=======================================
  Coverage   95.38%   95.38%           
=======================================
  Files         109      109           
  Lines        4182     4182           
=======================================
  Hits         3989     3989           
  Misses        193      193           

Flag	Coverage Δ
unittests	95.38% <100.00%> (ø)

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

Impacted Files	Coverage Δ
doctr/transforms/functional/pytorch.py	100.00% <100.00%> (ø)
doctr/transforms/functional/tensorflow.py	100.00% <100.00%> (ø)
doctr/transforms/modules/base.py	91.22% <100.00%> (ø)
doctr/utils/geometry.py	98.78% <100.00%> (+0.09%)	⬆️
...dels/detection/differentiable_binarization/base.py	91.19% <0.00%> (-0.63%)	⬇️
doctr/models/builder.py	93.75% <0.00%> (+1.04%)	⬆️

---

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 323d484...13e6b3b. Read the comment docs.

[charlesmindee]

thanks!