How can we statistics mAP after MGP?

[lishangyu]

dear author,
Thank you did a great job! I have thought about using optical flow to suppress false negative on object detection mission.But I am confused with mAP calculation, chould you tell me the formulation about mAP?
When I use caffe run some detection code like SSD,fast-RCNN,it chould calculate mAP when the forward propagation finished, I think MGP is a idea after CNN-net,so how can we statistics mAP again?Meanwhile, did this algorithm improve either mAP or recall. I thought is should be recall.
thanks a lot!!!!!

[wudiX]

I meet with the same problem as yours. I'm trying to understand the the code calculating the mAP in SSD layer 'Detection_output_layer'.
By the way, do you really understand the principle of MCS (mult-context suppression) in this paper. It says the MCS is for suppressing those false positives. It first sort detection scores in descending order. Then it divides scores into 'high-confidence' and 'low-confidence', and those 'low-confidence' scores are suppressed. But I think that if the score of false positive is very high, the false positive will be considered as 'high-confidence'. This is not we desire. So is MCS really works in improving detection accuracy?