tlpss
commented
8 months ago Hi @brambozz, thanks for the feedback!
Not sure if this matters in practice though, as predictions maybe never will have exactly the same probability :)
Yes I believe the impact is limited as this won't occur (often).
When points have the same probability they will still be ordered randomly, whereas they should be processed as a group and give 1 precision/recall value.
I had never thought of this edge case actually. But what you are saying makes sense! I would have to find another implementation of mAP calculation and see how they are handling this (e.g. Coco). If you dig deeper into this, feel free to let me know what you find! Otherwise I will keep it in mind and look into it when I have some time :+1:
brambozz
Hi!
I'm working on a 2D detection problem (bubble detection in ultrasound) and was looking for an AP implementation which thresholds on distance instead of IoU. Your implementation was exactly that, so thanks!
I noticed if I generate predictions with the same probability, I actually get a PR curve with many points, where I expected only one point. The source is I think this line, where points are ordered according to probability. When points have the same probability they will still be ordered randomly, whereas they should be processed as a group and give 1 precision/recall value.
Not sure if this matters in practice though, as predictions maybe never will have exactly the same probability :)