Unwanted 180-Degree Rotation and Text Recognition Errors in General Table Recognition V2 Pipeline

[devenus1]

I'm using the "General Table Recognition V2 Pipeline," and I've encountered an issue where the pipeline occasionally rotates vertically correct oriented images by 180 degrees before performing text detection and recognition. As a result, when the text is extracted, PaddleOCR fails to recognize the text properly, outputting random or gibberish strings like "eeeeer e, tyter."
Request for Help:
Can you help me identify why the pipeline is performing this unintended 180-degree rotation and provide suggestions for how to prevent or correct this behavior?

Code Used

from paddleocr import TableRecognitionPipelineV2

pipeline = TableRecognitionPipelineV2(
    use_doc_orientation_classify = False,
)
output = pipeline.predict(
    input = "processed_images/page_3.png",
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
)
for res in output:
    res.save_to_img("./output_ch/")
    res.save_to_xlsx("./output_ch/")
    res.save_to_html("./output_ch/")
    res.save_to_json("./output_ch/")

[Bobholamovic]

Since you set use_doc_orientation_classify to False, the auto-rotation feature should be disabled. Are you sure it's the whole image that's being rotated before being sent the the text detection network, and not the text line sub-images?

[devenus1]

The output saved after processing is the inverted image.
res.save_to_img("./output_ch/")

[liu-jiaxuan]

Please try the following settings:

output = pipeline.predict(
input = "processed_images/page_3.png",
use_table_orientation_classify=False,
use_doc_orientation_classify=False,
use_doc_unwarping=False,
)

[devenus1]

Thanks, it stops the rotation.

[bit-scientist]

I faced the exact same issue which was resolved by use_table_orientation_classify=False into the predict pipeline.

@liu-jiaxuan, could you explain why this happens with only some pages, but not with all? Don't you think it must be disabled by default?