Restore remap_boxes function

[Rob192]

Bug description

Following #731, the function remap_boxes and its associated test test_remap_boxes that were introduced in #488 were removed. This prevents the correct remapping of the boxes after they were rotated (parameter straighten_pages of OCRPredictor set to True).

Code snippet to reproduce the bug

from pathlib import Path

import matplotlib.pyplot as plt

from doctr.io import DocumentFile
from doctr.models.detection.zoo import detection_predictor
from doctr.models.predictor import OCRPredictor
from doctr.models.recognition.zoo import recognition_predictor
from doctr.utils.visualization import visualize_page

file = Path(r"myfile.jpg")

doc = DocumentFile.from_images(file)
det_bs=2
reco_bs=128

det_predictor = detection_predictor('db_resnet50', pretrained=True, batch_size=det_bs, assume_straight_pages=True)
reco_predictor = recognition_predictor('crnn_vgg16_bn', pretrained=True, batch_size=reco_bs)

predictor = OCRPredictor(det_predictor=det_predictor, reco_predictor=reco_predictor, assume_straight_pages=True, straighten_pages=True)

# Plot OCR output
out = predictor(doc)

fig = visualize_page(out.pages[0].export(), doc[0], interactive=False)
plt.show()

Error traceback

No error, however, the boxes that appear are misplaced.

Environment

Note: I pulled the last version from the repo but the DocTR version still appears as 0.4.1a0.

DocTR version: 0.4.1a0
TensorFlow version: 2.5.0
PyTorch version: 1.9.1+cpu (torchvision 0.10.1+cpu)
OpenCV version: 4.5.2
OS: Could not collect
Python version: 3.8.7
Is CUDA available (TensorFlow): No
Is CUDA available (PyTorch): No
CUDA runtime version: 11.0.221
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.0\bin\cudnn_ops_train64_8.dll

[fg-mindee]

Hi @Rob192,

Thanks for opening the issue 🙏
So I suggest doing the following in a PR:

• starting by designing a relevant unittest where we mock a white page with only a few lines (easy)
• we rotate this page, and we create a unittest where passing this image to estimate_orientation should yield the rotation angle with some margin (I think this is already setup)
• now we need to add almost the same test, with either light random or pretrained predictors with the straighten_pages option.

This should produce an error according to what you say. So then, in the same PR, we try to fix this so that it works 👍
What do you think @Rob192? 🙏

[Rob192]

Hello @fg-mindee
Sounds good !

[fg-mindee]

Alright! I won't be able to take a look at this before at least next week, but I consider this to be mandatory for the v0.5.1 patch release. @Rob192 Do you think you could work on PR in the coming days? 🙏

No problem if you can't, we'll postpone the patch release in that case!

[Rob192]

Hi, yes I can try to find time for this PR this week !

[Rob192]

Hi @fg-mindee, I was busy these past days, I could not get back to you earlier. I pushed the PR #812 to fix this issue !

[Rob192]

adding images as requested by @charlesmindee
WITHOUT remap_boxes (current)

WITH remap_boxes (#812)