feat: add filter_boxes method

[charlesmindee]

This PR adds a filter_boxes method in the core DetectionPostProcessor which performs the following:

• NMS with a high threshold to eliminate boxes included in others. NMS is implemented in Numpy (Fast-RCNN implementation).
• Merging for boxes with a consequent IOU, to avoid overlapping of letters which would appears in both words.

For now, only straight bounding boxes are supported, as the computation of NMS/boxes to merge is slightly harder when dealing with rotated boxes. In case of rotated boxes, it will simply return the array of boxes without filtering.

Any feeback is welcome!

[codecov]

Codecov Report

Merging #332 (400315c) into main (9bc289b) will increase coverage by 0.03%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #332      +/-   ##
==========================================
+ Coverage   94.69%   94.72%   +0.03%     
==========================================
  Files          83       83              
  Lines        3296     3335      +39     
==========================================
+ Hits         3121     3159      +38     
- Misses        175      176       +1     

Flag	Coverage Δ
unittests	94.72% <100.00%> (+0.03%)	⬆️

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

Impacted Files	Coverage Δ
doctr/documents/elements.py	98.93% <100.00%> (ø)
doctr/utils/geometry.py	100.00% <100.00%> (ø)
doctr/utils/metrics.py	96.59% <100.00%> (+0.87%)	⬆️
...dels/detection/differentiable_binarization/base.py	91.08% <0.00%> (-0.64%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9bc289b...400315c. Read the comment docs.

[fg-mindee]

Thanks for the PR! I added a few improvement suggestions

[fg-mindee]

Also, this PR modifies the default output behaviour (cf the failed unittest in the API)

[charlesmindee]

Changes:

• nms is now implemented in metrics, but not used here.
• the filter_boxes method uses box_ioa (which is implemented in metrics) to merge boxes above a threshold. This will remove boxes included in others or merge those who overlap significantly.
• a resolve_enclosing_boxarray function is implemented in geometry.
• added corresponding tests.

[fg-mindee]

Nice, thanks!

For further reference, since this changes the default behaviour of the detection part, should we update the benchmark tables as well? (to ensure it's a positive change)

[fg-mindee]

Sorry, I had another thought of merging both bbox resolution function since they indeed have the same purpose. The rest looks good to me!

[fg-mindee]

Thanks a lot!