It‘s still a bit flaky getting video acceleration (not talking about object detection but video decoding) working but after that it is one of the best solutions for live object detection I‘ve ever tried: no more small animals waking me up in the night.
P.S.: I‘m also supporting them with a yearly? subsciption to train the „A.I.“ model against false positives I provide which increased the accuracy even more.
This is becoming a real problem because the drivers/software for the Coral AI boards is yet another example of Google Abandonware(tm) which has a hard dependency on a Paleolithic-era version of Python. Comically, the hardware is still sold.
In so many words if you expect to use the Coral boards you are stuck on EOL versions of Debian/Ubuntu - which have terribly old video drivers and missing kernel GPU support. There's a good chance your modern GPU - even well-supported Intel ones - won't work.
Imagine buying new hardware in 2025 whose software still required Windows 7.
It's fine, you're not running a network-accessible part of the service on unpatched software. The only input this part of the software requires is trusted configuration data and a video feed which could hypothetically be malicious, but then the question becomes why you're running an adversarial camera on your network, and why you're allowing it to connect to the internet to fetch latest exploits and C&C instructions.
You can also transcode the video before feeding it to any outdated software and run it in a VM if you're paranoid.
Yes, it is, because then you aren’t stuck with a EOL distribution where you get even more security issues to deal with (vs. just EOL Python).
Also, what kind of “security exploits” would an outdated Python result in if the Python interpreter itself isn’t serving a network port or accepting arbitrary user input in general?
I assume Frigate itself isn’t running the web app on the same Python version - it’s likely just the Coral SDK that requires an outdated Python version.
Hedgehogs are fantastic TV - a member of my family used to get some great footage including one very memorable fight where one ended up rolling the other one around
Some nights my cat goes absolutely ballistic running from window to window to door, meowing and scratching to get out. And inevitably if I open my camera and look I'll see something like a family of racoons walking by or a skunk in the yard. It's a little consolation that he's just hearing other animals and isn't possessed by demons at 2am.
You can do both. You can set it to detect animals but turn off reviews. The reviews act like alerts you can "view" whereas the detection as more like metadata you can use on the search page
Been running about 2-3 years, was mostly fine before but now I get constant false positives from the children's garden toys, scooter left in the garden, pirate flag waving etc.
I don't submit false positives for privacy reasons but I'm looking at trainingy own model. I've got years worth of positives/negatives to train on.
That "subscription" is one which I gladly pay due to multiple reasons:
1. It supports the developers(s)
2. The price can be directly attributed to cost for training
3. You can keep the models you trained during your subscription indefinately
That's pretty much the opposite to AgentDVR. I don't need hosted services for remote access or push notifications - I can do that myself. But if I want to abide the license terms, I need to purchase a monthly subscription for remote access over my own VPN.
P.S.: I‘m also supporting them with a yearly? subsciption to train the „A.I.“ model against false positives I provide which increased the accuracy even more.