Metrics

[brudfors]

• I think BinaryPairwiseMeasures.youden_index is incorrect, fix: https://github.com/csudre/MetricsReloaded/blob/f7f6296468b6556c882cccebb8b6a973d2352e01/pairwise_measures.py#L273
• Is MultiClassPairwiseMeasures.weighted_cohens_kappa correct? I think perhaps the numerator should be divided by the number of observations? (from looking at "Weighted kappa: Nominal scale agreement provision for scaled disagreement or partial credit"): https://github.com/csudre/MetricsReloaded/blob/f7f6296468b6556c882cccebb8b6a973d2352e01/pairwise_measures.py#L129
• The reference I found for MultiClassPairwiseMeasures.matthews_correlation_coefficient uses the confussion matrix: https://en.wikipedia.org/w/index.php?title=Phi_coefficient&oldid=1105990453#Multiclass_case, not meausured if covariance. Maybe a reference to the current implemenation could be included for comparing?
• To correspond with paper, I think SQ in MixedLocSegPairwiseMeasure.panoptic_quality should be divided with TP: https://github.com/csudre/MetricsReloaded/blob/f7f6296468b6556c882cccebb8b6a973d2352e01/mixed_measures_processes.py#L66
• I have modified the metrics that depend on Trapezoidal integration because I was not sure the integration implementation was correct. The integration can be moved to a comon utility library like: https://github.com/csudre/MetricsReloaded/blob/f7f6296468b6556c882cccebb8b6a973d2352e01/utils.py#L188 and used when needed, e.g. https://github.com/csudre/MetricsReloaded/blob/f7f6296468b6556c882cccebb8b6a973d2352e01/prob_pairwise_measures.py#L192
• Could the Dice score implementation: https://github.com/csudre/MetricsReloaded/blob/64ac247dbc685d3e222740fe7a19aef403245e8c/pairwise_measures.py#L558 use fbeta?

[csudre]

Regarding segmentation_quality, SQ considers the average of iou for TP elements (so division by number of TP is already included)

[csudre]

Thanks for flagging the Youden Index issue - fixed now