Custom AI training
Hi Ben,
I was wondering if there will be a point where we can do our own AI training for custom things. My example is a pesky tree that when the wind blows triggers the motion detection. I can mask it out, but theat creates a lot of missed other motion. Shadows from other trees is another example.
I ask because those are my problems :) and would be a custom training vs car or animal things that benefit the whole system.
Clearly a long shot, but thought I would put it out there.
Eric
Comments
-
Not currently possible with SecuritySpy, but you will be able to train custom models soon with software called Frigate with their Frigate+ service.
I use SecuritySpy as my NVR, but do object detection through Frigate since it's much faster thanks to Google Coral support. It classifies objects in 8ms as opposed to 55ms in SecuritySpy.
Here is info on Frigate+
It's not a replacement for SecuritySpy since it's nowhere near as good as an NVR, but it does complement SecuritySpy quite well, at least in my use case.
-
Interesting and thank you
-
Hi @Mr_C thanks for the suggestion, we would like to offer something like this but it's not a trivial thing to implement. Probably in the near term, an intermediate step to this could be for users to review previous detections in some nice interface in the software, which would then send these results to our server, which we could then add to our database to use for further training. I think this could work quite well, and would help to improve the AI algorithms for all users.
@emergent - classification time in SecuritySpy is highly dependent on the hardware you are running on. On an M1 Mac, SecuritySpy can classify frames in 2ms.
-
@Ben - I agree. I was thinking the only way this would work would be through some sort of user based training (your near term suggestion). What I did not know is if you would then leave it as a local model or if it would go back to your database for the greater good. My example is a stupid Palm tree that was planted long after the camera was there. So if you had that ability, I would be training the system to recognize and then ignore the tree. Not super helpful for anyone without a Palm tree :)
But the overall concept might help you improve other aspects. Lots of us saying "yep that is a dog", etc.
Food for thought. Probably not the highest priority.
Eric
-
Hi Eric,
It's certainly a good idea when done correctly. I think the best system would send images back to our servers, where they would add to the pool of data used to train the next version of the algorithm. If your palm tree is causing you problems it's likely that a similar situation is causing problems for other users too - the wider variety of images we have in our training dataset the better.
For now, the best thing you can do to improve the algorithm is allow your instance of SecuritySpy to send us images from this camera (and any other you choose) using this option in the General Preferences.
-
Yes and thank you. I found that and turned it on.
