How to create precision recall curve using MeanAveragePrecision

[kylekam]

Hello,

I'm looking to create precision recall curves per class using the MeanAveragePrecision API. Precision is extracted easily enough using something like the following, but is recall missing a dimension? My understanding is that the PR curve is generated by calculating the precision-recall pairs while varying the confidence threshold [0.0-1.0].

p = metrics.compute()["precision"]
p_pairs = p[iou_idx, :, :, 0, 2] # extracts [num_pr_pairs, num_classes]
r = metrics.compute()["recall"] 
r_pairs = ... # missing the R dimension

My best guess right now is that recall_thresholds should be used in PR pairs and some form of interpolation is occurring instead of recall being explicitly computed.