Feature Request: Ability to Disregard Insects for AI Animal Detection
Forgive me if this was already requested or discussed elsewhere on the forum, but I didn’t find anything after a quick search.
I love the AI motion detection in SS, especially the animal detection. It’s proven to be much more robust than HomeKit, to the point where I’ll leave SS animal motion detection on in addition to HomeKit detection to make sure I don’t miss anything interesting. We live on the edge of a state park, so there’s a lot of animal life around. Unfortunately, that also includes insects, which leads me to my question.
Are there any plans to allow the exclusion of certain types of nuisance animals, like insects, from AI motion detection? If there are no plans, is this something that could be considered for the feature roadmap?
Comments
what insects are triggering the animal AI detection? Do you have your motion detection really sensitive? Are that insects that are close to the camera? Only small birds seem to trigger animal detection, never insects.
The motion events for insects are triggered primarily by wasps and bees. Motion detection is set to trigger after 0.5 seconds. Anything higher than that misses most events, including those that I’d like to capture. It’s only triggered by insects close to the camera, nothing far away.
When you say that only smaller birds seem to trigger animal detection and never insects, are you speaking from your personal experience? My experience is not the same; insects frequently trigger animal detection.
Yeah from my personal experience. They prob look like small birds to ur camera if theyre big enough insects close to the camera
have you tried adjusting these?
As mentioned, increasing the trigger time will reduce false-positives, but it will also increase false-negatives (missed events), so there is a tradeoff here.
SecuritySpy's animal detection is very accurate, but it's not quite as accurate as its human and vehicle detection. This is for a few reasons that are somewhat difficult to overcome:
We are actively working to increase the number of animals and bugs in our training dataset, so you can expect to see the accuracy of the animal detector increase over time!
Thanks for all the feedback. Just to clarify my request, I think animal recognition is great and works well. Sure, it occasionally identifies a wasp as a person, but that’s expected and I understand it’ll improve as the models are improved with new training data.
My request is to add more granular options for selecting what does and does not trigger motion detection, specifically the ability to remove insects from triggering motion. This would allow me to keep the animals option enabled so I could see most other animals while removing the “noise” of the flying insects. I understand I’d still get false positives, but 10 false positives a day versus the 100+ insect fly-bys I currently get would be a huge benefit.
you didnt give much info about your camera and settings.
how many megapixels is your camera?
whats its field of view like?
what kind of animals is SS bot triggering with?
is it because the animals are to fast, to slow or is the cameras focal length bad?
the insects that trigger, is it all day? Mostly at night? In InfraRed at night? In very lit up environment at night?
yea software can always be improved, but maybe you can improve your camera/environment/settings to get what your looking for.
Thanks for the continued feedback, although I'm not sure how changing my environment can get the results I'm looking for. In my case, SS is accurately identifying insects that fly in front of my cameras as "animals" and triggers motion events when that occurs. Since I've configured SS to trigger motion events for animals, this is the expected behavior. No changes are needed with that. The change I need is for SS to have the option to stop triggering motion alerts when it detects an insect. For example, it would be great if the "animals" category were broken down into "insects" and "not insects" sub-categories (not literally, just for the sake of illustrating my point). In that case, I'd select "not insects" and deselect "insects" to get what I want.
Insects arent suppose to trigger animal detection is the thing. So the AI is not working well enough in this case.
@killerdan56 is correct - insects are not meant to trigger animal detection. Unfortunately, flying insects can look like birds sometimes and this can cause a trigger (because we do classify birds as animals). We're actively working to add more animal and insect images to the training dataset in order to improve the accuracy of this.
@kappa12 it would be interesting if you could perhaps post some images of the insect captures you are seeing. This would help us understand your setup better, and also how the AI could be giving the wrong result here.
Oh! I didn't realize insects weren't supposed to be classified as animals in SS. In that case, now I understand the feedback I've received 🙂 Sorry that I wasn't aware of that, I could have saved us some back and forth.
I've attached a few snippets of insects captures that were classified as "animals". It's mostly the same insect (a mud dauber wasp) that causes the large majority of events. Sorry for the low quality, but I don't currently keep motion capture snapshots due to the high volume of events. I grabbed these images from the push notification previews.
Damn that things huge lol
WHAT ARE YOUR SETTINGS AT?
you prob need to save some footage with specific settings so you can send it to ben.
Wow yes that wasp is huge in the camera's view! It's most likely looking like a bird to the AI.
I'm wondering if perhaps we should remove birds from our animal training dataset so that they are ignored. Or, at least, have a separate category for them so that they can be filtered out. Many users won't want to include birds, and the downside of us including them in the animal category is that it makes it much more likely for flying insects to be classified as animals. Please let me know your feedback on this suggestion. This may need a bit more thought.
What im wondering is what animals is the AI not triggering for? Is it ones that you havent trained the your animal AI for.
Some animals (mainly insects) we have specifically trained the AI to ignore, because they are pretty much always a source of false-positive detections that users will want to ignore. Other animals are simply missing from our dataset entirely because they are not common in CCTV footage (snakes, elephants, moles etc.), so the detection of these cannot be predicted.
The animals for which we have most images in our dataset (dogs, cats, foxes, racoons, horses, deer, hedgehogs, rats, birds) are the ones for which the detection is going to be the most accurate.
I think what we'll do is create a separate "bird" category so that users can choose to include or exclude birds in the animal detection. Excluding birds will have the benefit of significantly reducing false-positive detections due to flying bugs.
@kappa12 - one more thing, if you haven't already, it would help if you allow the sharing of images with us from this camera. You can turn this on in the General preferences. It would be great to add some of the images of this beast to our training dataset. Thanks.
These are the settings I use for all of my outdoor cameras that are set up for AI-based motion detection:
Just a note that I have the system-wide thumbnail crop set to pretty tight, so since the images I uploaded before were from notifications on my iPhone, the wasp looked massive. After I uploaded those pics I enabled snapshots for motion events so I could get the whole frame and give better perspective on the size of the insect in the overall frame. I've included some examples below.
FWIW, I like the idea of creating a "bird" category that can be excluded based on the user's preference. I think this would go a long way towards solving the specific problem that I'm having.
@Ben , I'm afraid my family is too privacy-conscious to share images from our cameras. It's not that we don't trust you, we just generally shy away from giving that kind of access to our family's day-today activities. However, I'm happy to manually send you pics that I know don't include anything that we consider sensitive (like the photos I've included below). Let me know if that would help.
Something that could be of interest is the kind of camera I use outdoors, which is the Axis M2026-LE Mk II. They're pretty budget cameras as far as Axis is concerned, but as soon as I powered these up I noticed that they're extremely sensitive devices. They capture dust floating in the air when IR is active, and birds that it captures look like they're hovering in front of the camera with no motion blur at all. I guess these things have pretty good shutter speed, which may allow capturing something as small as an insect and contribute to higher motion detection false-positives as "birds".
From the SS manual
"Maintenance - Send anonymous images back to developer
If this option is on, SecuritySpy will send images back to us when motion is detected. By analysing and categorising these images, we can use them to improve SecuritySpy's motion-detection features. The images are scaled down to a small size and are sent infrequently over an encrypted connection, so the transmission is secure and uses very little bandwidth. Apart from the content of the images themselves, no identifying data is sent along with these images."
maybe when its a buggy day constantly triggering you can try this(if updated for animals)
https://bensoftware.com/blog/optimising-ai-object-detection/
to see what kind of percent the bugs are getting classified as animals, but regardless I'm sure its because they having wings like birds that it triggers. So Ben removing birds from his animal model most likely your best fix.
The smallest animal SS has detected for me
Thanks for the info and images. I think the separate bird category will significantly help with this issue if not solve it completely, so we're going to move forward with this.
I just stumbled upon this and 100% love this to be done. My carport camera watching my car some nights can detect animals (the bugs around the camera lens) every 30 seconds or so. When it happens I have to manually override the alerts but alas now I am awake I can not go back to sleep. So love the idea! 😎
Hi everyone.
I thought I would throw some images into the mix as well. I am a professional Artist and One image I can kinda see the head of a bird.
I have just animals selected because I want to scare the deer away with triggering the sprinkler systems in the appropriate zone. The settings are 70 on the animals and 70 on the motion 1.5 sec for trigger - everything else is 50.
See if anyone else can see the animals ???
Dale
technically humans are animals :P
I'm not complain anything just want to share some of my animal detected photos. It's meet my expectation. Except one big problem that SS always label my wife as animal.
I was considering projecting these for halloween -
@Ben I never responded here after you released the option to exclude birds like we discussed above. Thanks for making that change! It worked well, until we started getting birds nesting in areas near our cameras and had to disable it to know where they're nesting and where to remove the nests. So here I am again in June with the same mega-wasp triggering alerts just like last year. I swear this bug has the memory of a goldfish, constantly checking and re-checking the same cameras, going in circles around my house for 6 hours a day...
So this has me thinking of other potential ways to solve my problem:
1) Allow customers to identify specific events to trigger on less frequently in the future. There could be an option on the notification labelled something like "Show fewer events with this content" that would automatically train an on-device model for what to ignore.
or
2) Allow customers to manually train their own models for events to be ignored. I have dozens of photos of this insect per day that I could manually upload to an on-device training model. This is a little less elegant than option 1 above, but potentially easier to implement?
I understand that implementing either of these options is no trivial task, and they'd each probably require a separate UI in the main SS Mac app to manage ignored events to remove/reset them. Even so, I wanted to throw it out there to discuss and also to see if it's something that's already being considered to help prevent false positives for any kind of unwanted event, not just this use-case.