Use opencv's dnn(trained with tensorflow) api could decrease detect score?

edit

# I use ssd_inception models trained single target's detection model

My env:

* tensorflow1.8-gpu

* cuda 9.0

* cudnn 7.3

* CPU:E5

tensorflow result:link text opencv result

* GPU:GTX 1080x2

* os:Ubuntu16.0.4

The result show

so I use tensorflow detect my image,I have get very good result and ervey target had been found(thr=0.4). the image like this: but I use opencv's dnn api detect the same image,the result is like this(thr=0.1): The question is:

* Can the code of tf_text_graph_ssd.py in opencv's github generator graph.pbtxt decrease detect accuracy(some import op was delete?)?

* How to get the same result with use tensorflow?

* Is tf_text_graph_ssd.py operator some optimize op?

more infomation please check