I am doing a project on deepstream sdk4 using Jetson Nano, and my goal is to first detect vehicle grills in input video frames, then detect plates inside the detected grill, and finally detect the characters inside the detected plate, following a hierarchical fashion. To do that, I use three different yolov3-tiny networks for grill, plate, and character detection, respectively. However I would like my character recognition network to work in asynchronous mode. I have noticed that by setting
this can be done, but the problem is that it seems like the asynchronous mode can only be used for classifiers and not detectors (meaning that it can only work if
), however my network is a detector which should localize characters inside the plate and draw bounding box around each character, or at least to have the detected characters along with their location in the metadata if they are not shown up on the video.
what can I do to run a detector (and not a classifier) in asynchronous mode?
please note that I already use klt tracker.