The computation of MAP in test.py actually is precision.
the computation of map
in test.py actually is recall.
@wangzhaolei Yes, you are right!
This should be corrected in 7dddf2a. See utils.utils.compute_ap and test.py.
The computation of mAP still seems incorrect. The right way is to compute AP for all instances of each object class separately, and then average AP score. But in the code the mAP is averaged for AP of each image. This would result in a much higher mAP score that the correct one.
What is the precision envelope in utils.utils.compute_ap ?
If the precision is sorted, then the precision envelope mpre is a concatenated array of np.max(precision)*np.ones(len(mpre)-1) and array([0])".
Most helpful comment
the computation of map
in test.py actually is recall.