Add pixel group and contour expand ops

[jeffreykuang]

Motivation

The pixel group and contour expand operation are usually used in text detection.
They are now in mmocr, which result in installation difficulty.
Put them in mmcv so that mmocr is compilation-free and can be installed more smoothly.

Modification

Add pixel group op and its unit test.
Add contour expand op and its unit test.

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

Merging #993 (ce5f416) into master (1a5bf76) will increase coverage by 0.14%.
The diff coverage is 100.00%.

❗ Current head ce5f416 differs from pull request most recent head aa2b4c5. Consider uploading reports for the commit aa2b4c5 to get more accurate results

@@            Coverage Diff             @@
##           master     #993      +/-   ##
==========================================
+ Coverage   64.61%   64.75%   +0.14%     
==========================================
  Files         152      154       +2     
  Lines        9817     9858      +41     
  Branches     1784     1791       +7     
==========================================
+ Hits         6343     6384      +41     
  Misses       3144     3144              
  Partials      330      330              

Flag	Coverage Δ
unittests	64.75% <100.00%> (+0.14%)	⬆️

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

Impacted Files	Coverage Δ
mmcv/ops/__init__.py	100.00% <100.00%> (ø)
mmcv/ops/contour_expand.py	100.00% <100.00%> (ø)
mmcv/ops/pixel_group.py	100.00% <100.00%> (ø)

---

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 1a5bf76...aa2b4c5. Read the comment docs.

[grimoire]

Hi
Variable names like q_n kernel_cv p_res are a little hard to understand, better use more meaningful ones.
And loops with a single assign statement could be replaced with std::transform, faster and more readable.

[grimoire]

About this is_edge flag.

The point does not need to be revisited if all direction has been labeled or out of bound. Is that right? Would it be better to avoid feeding these points to next_queue?

[jeffreykuang]

given a pixel, if its all neighbors are labelled or out of bound, it indicates it is on edge. Therefore, it is the start point to expanding.

[grimoire]

I still don't understand, for example:

label_map=[
[0,1,0],
[1,1,1],
[0,1,0],
]

label_map[1,1] is not on edge but it will be passed to next_queue in every loop cause text_line[tmp_x][tmp_y] > 0

[jeffreykuang]

(1,1) can not be in the queue. because the queue is initialized with edge pixels and re-initialized with queue_next.

[jeffreykuang]

(1,1) can not be in the queue. because the queue is initialized with pixels on edge and re-initialized with queue_next, which are again pixels on edge.

[jeffreykuang]

label_map can not be=[
[0,1,0],
[1,1,1],
[0,1,0],
]
It can be
label_map=[
[0,1,0],
[1,0,1],
[0,1,0],
]

[grimoire]

I see, thanks for the explanation.

[jeffreykuang]

Hi
Variable names like q_n kernel_cv p_res are a little hard to understand, better use more meaningful ones.
done
And loops with a single assign statement could be replaced with std::transform, faster and more readable.
done

[innerlee]

Kindly ping @hellock