Implement Adversarial Texture Attack on Object Trackers

[beat-buesser]

Description

Continuing from PR #1353

This pull request implements the Adversarial Texture attack on object trackers by Wiyatno and Xu, 2019 (https://arxiv.org/abs/1904.11042) in PyTorch compatible with ART's PyTorchGoturn object tracking estimator. This implementation supports segmentation masks of foreground objects to train adversarial textures that are robust against being partially covered by the tracked or other objects.

Fixes #1326

Type of change

Please check all relevant options.

• Improvement (non-breaking)
• Bug fix (non-breaking)
• New feature (non-breaking)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[codecov-commenter]

Codecov Report

Merging #1430 (0f23952) into dev_1.9.0 (3cdd1ae) will increase coverage by 8.96%.
The diff coverage is 89.09%.

@@              Coverage Diff              @@
##           dev_1.9.0    #1430      +/-   ##
=============================================
+ Coverage      81.55%   90.52%   +8.96%     
=============================================
  Files            235      236       +1     
  Lines          19049    19214     +165     
  Branches        3365     3387      +22     
=============================================
+ Hits           15535    17393    +1858     
+ Misses          2807     1051    -1756     
- Partials         707      770      +63     

Impacted Files	Coverage Δ
...adversarial_texture/adversarial_texture_pytorch.py	89.02% <89.02%> (ø)
art/attacks/evasion/__init__.py	100.00% <100.00%> (ø)
art/estimators/classification/tensorflow.py	82.80% <0.00%> (+0.41%)	⬆️
art/estimators/classification/scikitlearn.py	90.86% <0.00%> (+0.66%)	⬆️
art/attacks/attack.py	92.59% <0.00%> (+0.92%)	⬆️
art/utils.py	90.92% <0.00%> (+1.03%)	⬆️
art/estimators/object_tracking/pytorch_goturn.py	76.05% <0.00%> (+2.11%)	⬆️
art/estimators/classification/keras.py	87.70% <0.00%> (+4.98%)	⬆️
art/attacks/evasion/iterative_method.py	100.00% <0.00%> (+9.09%)	⬆️
art/attacks/evasion/elastic_net.py	99.33% <0.00%> (+16.00%)	⬆️
... and 17 more

[lgtm-com]

This pull request introduces 9 alerts when merging fb2db39 into 451d909 - view on LGTM.com

new alerts:

• 9 for Module is imported more than once

[lgtm-com]

This pull request introduces 9 alerts when merging a27d478 into 96ad148 - view on LGTM.com

new alerts:

• 9 for Module is imported more than once

[lgtm-com]

This pull request introduces 9 alerts when merging cfc743a into 3cdd1ae - view on LGTM.com

new alerts:

• 9 for Module is imported more than once

[GiulioZizzo]

Hi Beat, many thanks for the review. I hope you find the comments helpful!

[GiulioZizzo]

size_average is deprecated, perhaps implement equivalent functionality using the reduction parameter

[beat-buesser]

Good point, I'll replace it with reduction='sum'.

[GiulioZizzo]

If a GPU is available, would it make sense to run the tests on a GPU to ensure that there is device consistency on this and future modifications of the code?

[beat-buesser]

That's good idea and would be very nice, at the moment there are no GPUs available on Actions. At the moment we run the tests manually on a GPU.

[lgtm-com]

This pull request introduces 9 alerts when merging 0f23952 into 3cdd1ae - view on LGTM.com

new alerts:

• 9 for Module is imported more than once