TF change antialias to true

[felixT2K]

This PR:

• changes antialias also for TF to True

NOTE tests:

• super minimal difference (< 1e-6) in the expected output (e.g.: 5.00001 instead of 5.0 / 1.00002 instead of 1.0) -
  I don't think there's much we can do about that (the results are correct and close to PT see below)

[felixT2K]

10 x abs
pattern: {loading}_{framework}_{loading}_{framework}_diff.jpg

both with antialiasing:

opencv_tf_pil_tf_diff:  108826.1
opencv_torch_pil_torch_diff:  1859450
opencv_tf_opencv_torch_diff:  6922.742
pil_torch_pil_tf_diff:  7202.461

only pt with antialiasing:

opencv_tf_pil_tf_diff:  428563.7
opencv_torch_pil_torch_diff:  1859450
opencv_tf_opencv_torch_diff:  470293.66
pil_torch_pil_tf_diff:  183107.75

[felixT2K]

observation over 30 images:

straight pages: no visible changes
rotated pages: narrow text is detected somewhat less well (should be fixed if we do the next run with rotated augmentation - TF #1364 )

@charlesmindee @odulcy-mindee wdyt ?

failing test will be fixed by #1336

[odulcy-mindee]

@felixdittrich92 I'm not sure to understand the benchmark. Why do you compare tf and torch with PIL and OpenCV ? Is there also antialiasing enabled by default in these frameworks when you downsize an image ?

[felixT2K]

@felixdittrich92 I'm not sure to understand the benchmark. Why do you compare tf and torch with PIL and OpenCV ? Is there also antialiasing enabled by default in these frameworks when you downsize an image ?

I compare the way how we load the image and the framwork which performs the resizing with and without antialising
After this PR we use in TF and PT antialias so the images are now much more close also for training runs

For more information: #402

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (e645ead) 95.76% compared to head (d02c4fb) 95.77%.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1348   +/-   ##
=======================================
  Coverage   95.76%   95.77%           
=======================================
  Files         155      155           
  Lines        6950     6951    +1     
=======================================
+ Hits         6656     6657    +1     
  Misses        294      294           

Flag	Coverage Δ
unittests	95.77% <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 Sentry.
📢 Have feedback on the report? Share it here.

[felixT2K]

Add the end what we want is to work on the same images in both frameworks in the whole pipe inference and training
and with antialias it's the closest we can get atm (in pytorch we use antialising already since 0.7.0)

i have compared the "loading" libs and processing frameworks to see the difference :)

The black images shows the smallest diff which are:

opencv_tf_opencv_torch_diff:  6922.742
pil_torch_pil_tf_diff:  7202.461

both with antialias=True

[odulcy-mindee]

Thanks !