[orientation] page orientation improvements

[felixdittrich92]

This PR:

• refactor OCRPredictor logic (limit duplication)
• extend estimate angle function to handle also upside-down images
• correct return value from estimate angle (negative - left side rot / positive - right side rot) -> update tests
• integrate page_orientation_predictor

NOTE: Draft until TF orientation models are merged

Any feedback is welcome 🤗

Closes: #1460

[codecov]

Codecov Report

Attention: Patch coverage is 96.96970% with 2 lines in your changes missing coverage. Please review.

Project coverage is 96.41%. Comparing base (b9c4e6d) to head (e9946de).
Report is 3 commits behind head on main.

Files	Patch %	Lines
doctr/models/_utils.py	91.30%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1553      +/-   ##
==========================================
- Coverage   96.44%   96.41%   -0.04%     
==========================================
  Files         164      164              
  Lines        7743     7773      +30     
==========================================
+ Hits         7468     7494      +26     
- Misses        275      279       +4     

Flag	Coverage Δ
unittests	96.41% <96.96%> (-0.04%)	⬇️

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

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

[felixdittrich92]

Maybe for better understanding:

To test I took some different images and rotated randomly in range (-180 - 180)
Example: (The test data contains much more dirty samples as the attached one 😅)

some results:

Page 0
orientation: 10, rnd: 11
Page 1
orientation: -106, rnd: -106
Page 2
orientation: -95, rnd: -96
Page 3
orientation: -132, rnd: -133
Page 4
orientation: 148, rnd: 147
Page 5
orientation: -99, rnd: -100
Page 6
orientation: 138, rnd: -42
Page 7
orientation: -118, rnd: -117

[odulcy-mindee]

I don't understand entirely this section. Here, estimated_angle == 90, so according to your comment, it's wrong, so page_orientation is returned but it's also 90. It means there is no evident fallback to fix this issue ?

[felixdittrich92]

yeah the problem case is if estimated_angle is exact 90 or -90 it's mostly not correct (swapping - and + or complete wrong)

So we should check the page orientation also..

I think we can simplify this:

if abs(estimated_angle) == abs(page_orientation):
    return page_orientation

wdyt ?

[odulcy-mindee]

But what "mostly wrong" means here ? Based on your experimentations, if it's 90 or -90, it's completly wrong ?

[felixdittrich92]

In the most cases yes

[felixdittrich92]

comes from:

elif w / h < 1 / ratio_threshold_for_lines:  # if lines are vertical, substract 90 degree
            angles.append(angle - 90)

but i don't have a better solution for this yet :/

[felixdittrich92]

@odulcy-mindee

pushed a small update

tested again with a complex set of documents (mobile phone, receipts, etc) 30 samples

randomly rotated in range -180 - 180

it mostly varies between 22 - 25 correct from 30

before this integration / changes on the same set we have had 7 - 11 correct so as mentioned still some space for further improvements it's not perfect but already much better as before :)

[felixdittrich92]

ah.. if i lower the min_confidence we get ~27 - 29 correct xD trust the page orientation model xD

[odulcy-mindee]

ok, it sounds good for a "first" iteration. We can improve later in another PR

[felixdittrich92]

yep i think the same