bug in mAP calculation

[iTomxy]

I found a bug in your code of calculating mAP, which lies in the function get_AP, originating from the wrong construction of the recall difference list. I advise to change the final computing code of AP to AP = sum(l .* pl) / sum(l);. For reference codes and testing case, see:

• calcMapTopkMapTopkPreTopkRecLabel.m
• calc_hammingranking.py

[kunhe]

Hi, thanks for your attention to this work and to the details. I will check on it this week. Assuming it's indeed a bug, I'm wondering if you have numerical experiments on how much it affects results?

[iTomxy]

Not yet, but in the simple case I mentioned, it's a 0.558333 with 0.704167 discrepancy. Although it lies in the range given by the actual lower bound and upper bound, the computed upper bound is wrong.