The implementation of COCO metric in the tutorial

I’m following the mrcnn tutorial on the website but I got confused when I evaluated my dataset. To make the prediction results more “reasonable”, I raised the “box_score_thresh” in the source code. (Default value is too low, maybe 0.25, that there would be some false detections with such low scores sometimes.) However, the AP value should be lower than before when it was with a lower threshold? I can’t see why, shouldn’t the FP decrease therefore get a higher AP?
I really appreciate it if someone can figure it out. An abstract evaluation process is also welcome. Thanks a lot!