1

I have a general question regarding the mAP score used in measuring object detection system performance.

I understood how the AP score is calculated, by averaging precision over recall 0 to 1. And then we can compute mAP, by averaging AP score of different labels.

However, what I have been really confused, is that, it seems that mAP score is used to denote the "precision" of a model. Then what about the "recall" aspect? Note that generally speaking, when measuring the performance of a machine learning model, we need to report precision and recall at the same time, right? It seems that mAP can only cover the precision aspect of a model.

Am I missed anything here? Or mAP score, despite its name is derived from Precision, can indeed subsume both "precision" and "recall" and therefore become comprehensive enough?

lllllllllllll
  • 239
  • 1
  • 6
  • average precision is relative to recall, its says something about the precision recall give/take. – mshlis Aug 21 '19 at 20:27
  • @mshlis. Thanks for the clarification. Yes, that's what I can think about.. But is there any more "formal" description/statement such that "mAP is derived from both precision or recall"? – lllllllllllll Aug 22 '19 at 01:15
  • Look at F score – mshlis Aug 22 '19 at 01:15
  • @mshlis Yes, I understood the definition of F score, but what's the relation with mAP? – lllllllllllll Aug 22 '19 at 01:24
  • @mshlis OK, I think I found what I am looking for. What I can say is that "mAP derieves from both precision and recall values, by computing the area under the precision-recall curve for a given class." In that sense, yes, it surely subsumes both precision and recall values in a way. – lllllllllllll Aug 22 '19 at 01:26

0 Answers0