[transforms] Crop the whole crop_box region in crop_detection

[MohammadHijjawi97]

crop_detection projected the relative crop_box onto (w - 1, h - 1) instead of (w, h). The image crop therefore lost its last row and column and no longer matched the region the boxes are rescaled to: relative boxes are divided by the crop's real size, while the image came out one pixel smaller and slightly shifted. Absolute boxes were also clipped to w - 1.

import numpy as np, torch
from doctr.transforms import RandomCrop
from doctr.transforms.functional import crop_detection
from doctr.utils import Sample

img = torch.arange(24, dtype=torch.float32).reshape(1, 4, 6)
crop_detection(img, np.array([[0.0, 0.0, 1.0, 1.0]]), (0.0, 0.0, 1.0, 1.0))[0].shape  # (1, 3, 5)
crop_detection(img, np.array([[0.0, 0.0, 1.0, 1.0]]), (0.5, 0.5, 1.0, 1.0))[0].shape  # (1, 1, 3), expected (1, 2, 3)
RandomCrop(scale=(1, 1), ratio=(1, 1))(Sample(image=torch.rand(3, 1024, 1024), target=np.array([[0.1, 0.1, 0.9, 0.9]]))).image.shape
# (3, 1023, 1023)

With this PR they return (1, 4, 6), (1, 2, 3) and (3, 1024, 1024). RandomCrop is used in the detection and layout reference training scripts, so their random crops now line up with the boxes and masks.

Tests:

• test_crop_detection: the expected shapes were (3, 26, 37) for a 0.54 x 0.76 crop of a 50x50 image; they are now (3, 27, 38). Added exact-pixel checks for a full crop and a quarter crop, with relative and absolute boxes.
• test_random_crop: a full-size RandomCrop keeps the image, the mask and the relative target unchanged.

Both fail on main and pass with the fix. ruff check, ruff format --check and mypy doctr/ pass.

Any feedback is welcome 🤗

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.62%. Comparing base (11941d2) to head (b1e0aa3).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2174   +/-   ##
=======================================
  Coverage   97.61%   97.62%           
=======================================
  Files         170      170           
  Lines       10740    10740           
=======================================
+ Hits        10484    10485    +1     
+ Misses        256      255    -1     

Flag	Coverage Δ
unittests	97.62% <100.00%> (+<0.01%)	⬆️

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

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[felixdittrich92]

Thanks @MohammadHijjawi97 👍