Agree. I clone the camera. One for old timey motion detection wit full time recording. The other for AI. Then, if necessary, I can look in the full time recording at the proper time.
No stupid questions.
What works for me, my targets may vary from yours.
Do not use Default Object Detection
Do not use Main stream if available
Use custom models.
I use 100 msec and 4 images, but that is mostly for moving vehicles. Trial and error called for here due to other variables. Like...
Running CPAI in a Debian VM, with Docker, lowly P620 Nvidia with 2GB RAM. In the CPAi log I see 51 msec, with 66 msec in BI log. That is typical, it varies. BI on W10 VM.
FWIIW: I am running via Docker in a Ubuntu VM, with Nvidia GPU via PCIE passthrough. One (minor) advantage is that I don't have CPAI installed on the BI systems. I have 2 test systems, one here in SOHO and one at a client's site. Client's site has 5 (of 11) cloned cameras facing street side and...
FWIIW: For testing, I am using AI via Docker in a Linux VM. I created a Samba share of the folder with the models and put that in the AI on BI:
\\192.168.50.17\cuda