Too many boxes when using demo.py with --video-input video.mp4

[BrainPotter]

Hello! Thanks for your great working!
I run demo.py with --video-input. But there are too many boxes in the generated results. Please tell me what I should do.

I Adjust the threshold value in the detector.py. This problem cannot be solved.

I also run demo.py with --input image.jpg, the results seems correct.

I really want to know why?
PS: trained on COCO and LVIS