Dogs constantly triggering "human detected"
Hi Ben, our small dogs trigger the "human detected" action despite having the sensitivity slider set to 90. Are there anything recommendations for diagnosing this, or any way we can help provide image frames to get the human detection algorithm improved to reject animals? Or is this expected unless at a level of 90? Thanks!
Comments
Dogs should be effectively screened out by the algorithm, so I'm curious to find out more here about what could be going wrong. As per the advice in the above blog post, once you have set up the "AI Predictions" folder, could you please email us some of the images that show the AI predictions for your dogs?
But the spider webs should not trigger the human AI too much. The most likely explanation for this is simply that our AI (like all such algorithms) is less than 100% accurate - in fact in our testing it is 95% accurate - so if you have spider webs moving continuously and triggering motion all day, then it will still let through 5% of these events, which may be a significant number of triggers. In this case, looking at clearing the spider webs, and perhaps improving the lighting (i.e. less light on the spider webs, more light on the area of interest), would be the solution here.
Just to follow up on this, we have now added an AI animal detector to the latest version of SecuritySpy (currently 5.5). This addition makes the AI much more able to tell the difference between animals and humans, so this should certainly help with the above problem. Any feedback is welcome.
holy cow! animal detection? this just changed my plans. thank you!!
Unfortunately my small dogs still trigger “human detected” constantly with 5.5. I see the confidence sliders have been defaulted back to 50. Should I bump these back up again higher to avoid dogs being recognized as humans? They look nothing like people in the trigger photos to my eye…
Hi @Mech33 In general, I wouldn't recommend adjusting the sliders - the higher you set them the more false-negatives you'll get (missed events).
In order to help us train the next update to the AI algorithm to better be able to distinguish between dogs and people, the best thing you can do is to allow sharing of the images from these cameras, so that we can add your images to our training dataset. There is an option for this in the General section of the Preferences in SecuritySpy.
@Ben are the “human identified” images of actual dogs being re-classified manually somewhere before training, or how are they being added to the “dogs not people” training dataset once automatically uploaded?
We are continuing to actively gather images of animals from our users, so any images that your system can provide will be useful. These are labelled by a human at our end, and then added to the training dataset for the next update of the algorithm. The more varied and numerous the images of animals we have, the better the accuracy will be.