has anyone tried doing object detection or face recognition on multiple camera feeds in near realtime (15fps)? Is deepstack able to handle it?
I can get up to 13 fps OD and 7 fps FR with deepstack running on 8GB Jetson Xavier. Surprisingly, that’s just slightly faster than the fps I get on 4GB Jetson Nano. It handles 3 parallel camera streams without much effort.
While 7 fps in FR is not realtime, I would assume it’s more than sufficient in most setups. Face recognitions are instantaneous. Even two-fold increase in fps performance would probably be unnoticeable. I’d suggest you implement a LIFO queue mechanism to provide deepstack with the most recent frame for processing. That’s practically realtime.
Do you have a sample script you can share that uses either a video stream or a video file?
Take a look at my ai gatekeeper script that controls my wicket gate. ipc.py and ds.py do exactly what you’re looking for.
@sickidolderivative the models are not optimised for the Xavier yet. Can be done with TensorRT
@robmarkcole One probably needs a PhD to do that
IMHO the beauty of Deepstack, besides performance, is the simplicity of use that makes AI accessible to non-technical people like myself.