[orientation classification] Result differs from docTR

[felixdittrich92]

Bug description

Ref.: mindee/doctr#1801

Earlier I had written a whole module which uses pdf2image to extract page image and run DocTR to find the orientation of the image. Later I realized that it's not the problem of DocTR but OnnxTR which I was using earlier.

I want to report a strange case where for the same image DocTR gives correct orientation of 0 as the image is vertical, while OnnxTR gives 180degree orientation.

Here's the image:

Code snippet to reproduce the bug

import os
import json
from doctr.io import DocumentFile
from doctr.models import ocr_predictor
# from onnxtr.io import DocumentFile
# from onnxtr.models import ocr_predictor

os.environ['USE_TORCH'] = '1'


img_path = "extracted_image/abc6_page_1.png"


predictor = ocr_predictor(
    det_arch="db_resnet50",
    reco_arch="crnn_mobilenet_v3_large",
    resolve_blocks=True,
    resolve_lines=True,
    paragraph_break=0.035,
    assume_straight_pages=False,
    detect_orientation=True,
)

doc = DocumentFile.from_images(img_path)
result = predictor(doc)
text_data = result.render()
json_data = result.export()
print(json_data)

parsed_file = f"doctr-output.txt"
with open(parsed_file, "w") as f:
    f.write(text_data)