In fact, in videotest.py, VideoTest.run() has a parameter conf_thresh, which is set to the expected value of 0.6, but then when calling detection_out(), it doesn't pass conf_thresh at all. Instead it just thresholds after the fact.
I've found that I can get substantial FPS improvements by passing the conf_thresh of 0.6 into detection_out() (usually around 1.5x increase)
EDIT: Actually that 1.5x increase is only for the default COCO model. For my own re-trained models on different datasets, I've gone from 2 FPS to around 25 FPS when using this fix.
In fact, in videotest.py, VideoTest.run() has a parameter conf_thresh, which is set to the expected value of 0.6, but then when calling detection_out(), it doesn't pass conf_thresh at all. Instead it just thresholds after the fact.
I've found that I can get substantial FPS improvements by passing the conf_thresh of 0.6 into detection_out() (usually around 1.5x increase)
EDIT: Actually that 1.5x increase is only for the default COCO model. For my own re-trained models on different datasets, I've gone from 2 FPS to around 25 FPS when using this fix.