I managed to run the back to back detectors with both models running on full frame. i.e. the secondary gie runs on full frame instead of bounding boxes from primary gie. But the output is slow. So I was thinking if we can select an ROI for the secondary to detect on. My primary is a hand detector and my secondary is a face detector. In this project, the face will not go below a certain point so there is no reason to detect it below that point. It might significantly improve performance. Any help in this regard is appreciated. Thank you.