attribution localisation metric should consider positive attributions only

[leanderweber]

Description

According to the paper (https://arxiv.org/pdf/1910.09840.pdf) attribution localisation should be computed only from the positive values in the attribution map. The implementation in quantus considers negative values also in the computation.

Minimum acceptance criteria

• [ ] Update implementation to only consider positive attribution values
• [ ] Update documentation of AttributionLocalisation to correctly state this
• [ ] Potentially update test cases, in case above changes break any of them
• [ ] I think the warning w.r.t. inside_attribution_ration > 1.0 can also be removed then
• review pls @annahedstroem

[annahedstroem]

The current implementation set abs=True

and the current docstring says

"""
Attribution Localization implements the ratio of positive attributions within the target to the overall
attribution. High scores are desired, as it means, that the positively attributed pixels belong to the
targeted object class.
"""

which for me is sufficient. Can we close this issue? @leanderweber