how to improve the model for real time tracker?

[chunniunai220ml]

Hi,
I have run the code on custom single video, and got satisfactory results, but it is time-consuming(about 2 fps for 720*720 frame on K80) for real time tracker.
So I wonna know how to improve the model for real time tracker?for example, using a detector like yolov3 (we should train the model from scratch?)
In addition, do you try add the const type=='person_reid' method in the _compute_distance_matrix for improve the track-id accuracy? if you have done this, can you tell me the effect？ if not, I plan to add the type to test the effect.