pliablepixels
commented
3 years ago I can't promise this anytime soon, but it falls in line with what I am doing with pyzm - allows you to arbitrarily combine multiple models/frames etc. It will take more work, but I can see this ask align with that approach, eventually. Today I allow arbitrary sequencing of models, but not underlying processors. I can add that. Even when I do, it will not be integrated to mlapi and zmES - that is more work for a future day.
Short answer: not anytime soon, but some day.
Hi, I've been using yolov4 model object detection with zm 1.35.10 and it is working well. Recently I came into possession of coral tpu. They both miss some persons in some events but I think it's because of the location of the camera, I rent the top floor flat so the cam looks down at walkway at almost 90 deg. angle. With that said it works better than I expected honestly.
I've set up mlapi server vm and usb-passthrough'd the coral tpu to it. Made sure all works, yolov4 detects same as with local object detection, coral tpu detects the bird from the installation instructions, and detects persons from zm too.
The thing is some events won't detect the person with yolo, some won't detect person with tpu. Since I have local hook processing set as fallback and it correctly falls back to local yolo from remote tpu if I turn off the mlapi vm, couldn't there be a setting to use both and either report the detection with higher %, or in case only 1 method reports back report that?
Just an idea. Many thanks for all your work I use and enjoy a lot of it!