Precision-Recall Curve Bugfix

[czaloom]

Issue

AP was inconsistent between v0.36.6 and v0.37.2 (latest) when tested on a larger real-world dataset

• Manual inspection of the computation cache yielded the following bugs
  • calculate_ranking_boundaries would sometimes include groundtruth-prediction pairs without matching labels in the iou_boundary array. This array should only include candidates for true positives.
  • For high-density clusters of annotations, the pr_curve binning operation within compute_counts would not necessarily return the maximum precision or score depending on ordering. When running over file-based chunks this would intermittently yield different results as there was more opportunity for the correct maximum value to be chosen.

Both of these bugs resulted in small errors that could accumulate into larger errors if using file-based chunking.

Changes

• updated calculate_ranking_boundaries to enforce label matching as a condition for inclusion
• updated compute_counts with a new binning operation.