Are the embedded TensorFlow model labels/percentages accessible?

AlaskanOffGridder
AlaskanOffGridder
in General

BLUF: it would be great if the actions logic could be based on AI model specific labels rather than categories in some way.

Only been evaluating for a week or so, but the AI is impressive. It has correctly recognized people, and ATVs and construction equipment as vehicles, and has recognized moose and squirrels as animals. Hasn't been tested on bears yet but that's coming any day now I'm sure.

From what I understand, the model is trained with TensorFlow and embedded into the software. I would imagine the model has the ability to provide animal types and an accuracy percentage. If that's the case, is this accessible in the software? It would be useful to not get emails for squirrels for example. But moose and bears we definitely want email alerts since that is useful before we pop out of the cabin and head to the garden...always good to know when the moose and bears are around. I have a couple of times walked out the front door and spooked a bear who was just outside the door or under the deck. But I literally have squirrels and birds on the deck constantly throughout the day and would like to keep from getting alerts for them.

Comments

  • Ben
    Ben

    Great to hear that SecuritySpy has been working well for you so far!

    At this time, the model outputs a percentage probability of an animal, but does not distinguish between different types of animal. This is because our model is trained on real CCTV images - this makes it very accurate in real-world situations, but the issue is that it's difficult to gather enough real CCTV images of lots of different types of animals, in order to train the model to accurately distinguish between them.

    We are gathering more data all the time, and this is perhaps something we could look at in the future. To help with this effort, please enable the option in the General Settings in the software to anonymously share images with us. It sounds like your animal activity would be a very useful and varied source of images for our training data!

  • AlaskanOffGridder
    AlaskanOffGridder

    Thanks for the info. Training on actual CCTV images makes a lot of sense. I may dabble with training a model on my end to analyze the footage that your software captures from our cameras. I can have a huge dataset of squirrels in no time flat and maybe I'll write a script to check the directory when your software saves a motion capture, then I can filter out notifications for squirrels. Also might try the same with human detection so that we don't get notifications about ourselves walking around.

  • jimmyjohnson
    jimmyjohnson

    @Ben if SS does not distinguish between different types of animals, then how do you filter out birds for the feature you created to reduce insect triggers?

  • Ben
    Ben

    Hi @jimmyjohnson you are correct - my statement above is an oversimplification. SecuritySpy does separately classify non-bird-animals, and bird-animals. If you have the "Animal detection should include birds" option enabled, then a hit on either class will cause an animal trigger. Otherwise, only a hit on the non-bird class will cause an animal trigger. This is the only way in which SecuritySpy distinguishes between different types of animals at this time.

  • jimmyjohnson
    jimmyjohnson

    Thanks for the info. Wasn't trying to "get ya" but rather wondering if there were deeper details that may be leveraged with scripting, even if not available via GUI.

  • Ben
    Ben

    Not at all - I'm glad you mentioned this as I did neglect to include this detail in my original description!

Categories