Improve mask_to_rle performance

[vinnamkim]

Signed-off-by: Kim, Vinnam vinnam.kim@intel.com

Summary

• Resolve [Performance] Converting mask to RLE comparison with COCO tools #745 and ticket no. 97760.
• Import a run-length encoding (RLE) code implemented in pycocotools and bind it with pybind11.
  https://github.com/cocodataset/cocoapi/blob/8c9bcc3cf640524c4c20a9c40e89cb6a2f2fa0e9/common/maskApi.c#L32-L41
• This new C-API will provide the uncompressed RLE same function as the previous API but shows the same performance as pycocotools.

import numpy as np
from datumaro.util.mask_tools import mask_to_rle, mask_to_rle_py
import pycocotools.mask as mask_utils
from time import time

img_size = 768
mask = np.zeros([img_size, img_size])
# mask = np.random.randint(0, 2, size=[img_size, img_size])

for i in range(img_size):
    mask[i, :i] = 1

mask = mask.astype(dtype=np.uint8)

t1 = time()
for _ in range(10):
    result = mask_to_rle_py(mask)
print(f"Datumaro old implementation: {(time() - t1) * 1000:.2f} (ms)")

t1 = time()
for _ in range(10):
    mask_utils.encode(np.asfortranarray(mask))
print(f"pycocotools compressed RLE: {(time() - t1) * 1000:.2f} (ms)")

t1 = time()
for _ in range(10):
    result = mask_to_rle(mask)
print(f"Datumaro C-API uncompressed RLE: {(time() - t1) * 1000:.2f} (ms)")

Datumaro old implementation: 25.32 (ms)
pycocotools compressed RLE: 10.40 (ms)
Datumaro C-API uncompressed RLE: 9.66 (ms)

How to test

If the new C-API is compatible with the previous API, our existing unit tests should be passed all.

Checklist

• I submit my changes into the develop branch
• I have added description of my changes into CHANGELOG
• I have updated the documentation accordingly
• I have added tests to cover my changes
• I have linked related issues

License

• I submit my code changes under the same MIT License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below)

# Copyright (C) 2021 Intel Corporation
#
# SPDX-License-Identifier: MIT

[codecov-commenter]

Codecov Report

Base: 84.61% // Head: 84.67% // Increases project coverage by +0.06% 🎉

Coverage data is based on head (f93dcd9) compared to base (e3ca3b5).
Patch coverage: 100.00% of modified lines in pull request are covered.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop     #770      +/-   ##
===========================================
+ Coverage    84.61%   84.67%   +0.06%     
===========================================
  Files          157      157              
  Lines        22055    22064       +9     
  Branches      4829     4829              
===========================================
+ Hits         18661    18683      +22     
+ Misses        2284     2273      -11     
+ Partials      1110     1108       -2     

Flag	Coverage Δ
macos-10.15_Python-3.10	84.58% <100.00%> (+0.49%)	⬆️
macos-10.15_Python-3.7	84.27% <100.00%> (+0.51%)	⬆️
macos-10.15_Python-3.8	84.51% <100.00%> (+0.49%)	⬆️
macos-10.15_Python-3.9	84.52% <100.00%> (+0.50%)	⬆️
ubuntu-20.04_Python-3.10	84.60% <100.00%> (+0.52%)	⬆️
ubuntu-20.04_Python-3.7	84.29% <100.00%> (+0.51%)	⬆️
ubuntu-20.04_Python-3.8	84.53% <100.00%> (+0.50%)	⬆️
ubuntu-20.04_Python-3.9	84.54% <100.00%> (+0.50%)	⬆️
windows-2019_Python-3.10	84.60% <100.00%> (+0.50%)	⬆️
windows-2019_Python-3.7	84.29% <100.00%> (+0.06%)	⬆️
windows-2019_Python-3.8	84.52% <100.00%> (+0.50%)	⬆️
windows-2019_Python-3.9	84.53% <100.00%> (+0.50%)	⬆️

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

Impacted Files	Coverage Δ
datumaro/plugins/synthetic_data/image_generator.py	93.49% <100.00%> (+10.66%)	⬆️
datumaro/util/mask_tools.py	76.78% <100.00%> (-3.83%)	⬇️
datumaro/plugins/synthetic_data/utils.py	84.61% <0.00%> (+2.56%)	⬆️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[bonhunko]

LGTM!