Deepstack found a bear in the front yard - I'm in trouble

Deeepstack analyzed this photo and identified my wife as a bear. Where do I send the bill for the flowers :slight_smile:

I get that she was bending over wearing a brown jacket and has long brown hair.


@sprior As hilarious as this detection is, I am sorry it may have given you a scare that a bear is wandering around your house. :joy:

What happened here is an example of an adversarial image, in which the AI model misunderstands the context of the image due to some unusual activity.

There is a chance the probability you got for that detection will be low relative to the average true positive person detections.

In the future, we will work on adding more resilience to the Detection API’s model to ensure it doesn’t get confused by such unusual images.

1 Like

I just thought that image and joke was too good not to share. And yes what I’d really like to do is to send an alert if a bear is detected so false positives would get people too nervous.

I do look forward to better rejection of adversarial images. My bigger problems lately is Deepstack detecting a car on my patio and people in my basement workshop, and that gets spooky.

The other thing on my todo list (which I tried to start another thread about but got nowhere) is to deal with items that are detected, then not detected, then detected again - often cars in my driveway. I’ve been brainstorming the concept that if I saw a car in a spot and then don’t see it, but then see it again a few minutes later it’s probably not a new car, just the old one. And what I’m interested in is notification of new objects. So I’m thinking about how to keep track of not only what was last seen, but an expiring idea of what was recently seen close to that spot in the picture. Then you get into the idea that some things (a parked car in a driveway) is expected to move less from frame to frame than a person would.

it’s not exactly the same, but have you looked into dynamic masking? I believe this can ‘cancel out’ objects that were already detected there, and allow you configure automated actions accordingly.

imo, no matter how good you get it, there is still the possibility it will miss something that the setup now deems a false positive. Maybe instead, for cars & humans, you could get a better camera setup that can actually ID these, by license plate or face recognition, or say custom training a model for your cars vs others, and then allow system to act on that information, which will be much more accurate than trying to guess if something was there or not / size of object / speed of movement / some other indirect guesswork.

Haven’t looked at that, will have to see what the concept is.

I understand that it’s always going to be an imperfect best effort, but I’m trying to work with a set of cameras that I’ve already got in place. Nothing is placed in a good spot to read license plates. I’ve got two different use cases in mind, the one I mentioned isn’t really trying to identify specific instances of an object type, just trying to get an input on whether a new one enters the field of view or leaves.

Of course the other use case is for some objects to identify the specific objects - our cars or a friends car, maybe specific people. But the two cases are different from each other.

Before I started using Deepstack I’ve been using Sighthound. What I don’t like about Sighthound is that it’s a remote service, but it does a pretty good job at object identification. Reading the license plate is only occasionally possible, but it does a good enough job at returning make/model/color/vehicle type that I can often decide that it’s my or my wife’s car, and it can identify a UPS/Fedex/Mail truck. If Deepstack could do that I’d be a very happy camper.

1 Like

One thing just occurred to me as a possibility to help with the detection/no detection oscillation problem. I’m currently using a 0.5 confidence threshold over that amount detects and below that doesn’t. Maybe once a higher confidence is used to detect an object in a certain location I should use a lower threshold to keep detecting that object in the same spot until the confidence drops below a certain threshold when I consider it gone and the higher threshold would be required to detect a new object there. Will have to play with that.

1 Like