A curious phenomenon in image retrieval

I do image retrieval task with --use_geometric_verification on the code perform_retrieval.py in delg folder. To compare different local feature, I tried to replace the local feature of delg with Superpoint. But the performance is lower than the baseline (do global retrieval only). as far as I am concerned, geometry verification is a commom method to improve the performance.

I wanna know that is this phenomenon normal? Does it means geometry verification can't improve the performance definitely?

the result of Superpoint is shown below:

hard
mAP=45.17
mP@k[ 1 5 10] [87.14 71.43 60.29]
mR@k[ 1 5 10] [19.29 29.39 36.42]
hard_after_gv
mAP=41.04
mP@k[ 1 5 10] [84.29 68.57 56.57]
mR@k[ 1 5 10] [19.01 27.18 31.21]
medium
mAP=69.75
mP@k[ 1 5 10] [95.71 92. 86.86]
mR@k[ 1 5 10] [10.17 25.94 33.84]
medium_after_gv
mAP=67.29
mP@k[ 1 5 10] [95.71 91.43 86.14]
mR@k[ 1 5 10] [10.19 25.67 33.6 ]