Dog / object AI detection
  • Any interest in updating the image detection to something like YOLOv3, which supposedly can detect 80 objects.

    YOLOv3
    https://arxiv.org/abs/1804.02767

    Article about it
    https://gizmodo.com/neighborhood-hero-builds-ai-powered-device-that-automat-1846562170
  • Better AI is going to be a crucial feature going forward. All the cloud providers are going to be able to do a lot in this area I feel. Hopefully those who process videos locally will be able to keep up.
  • Does this go between the camera and SS?
  • It's unlikely that we'd integrate such a package. For SecuritySpy, we need algorithms specifically trained on CCTV images in all lighting conditions, and that run on Apple's CoreML platform in real time. This is exactly how our own human and vehicle detectors have been designed, and they work very well. And, most users of CCTV software are really only concerned with detecting humans and vehicles - adding more classes of objects that the vast majority of users will never need will just result in worse overall performance for most users.
  • One main use case is dog pooping on private property, and the owner didn't pickup. Thus wondering if dog can be separately identified.
  • I would LOVE a way to ignore dogs. Our dogs trigger the human detection all the time, even when we set it for over 95%. And they’re miniature schnauzers… :)
  • would love to add an animal detector - dogs, raccoons, coyotes, foxes, etc.
  • I also would love to see an animal detector. One of the things we liked about BI was seeing what critters walked in front of our cameras, usually at night. Cats, raccoons, coyotes, possums, foxes, deer and turkeys have wandered by over the years.
  • There are a number of problems implementing an animal detector:

    - Animals come in all shapes and sizes. What should qualify as an animal? A bird? Fish? Insect? Different users will have different opinions.

    - It is very difficult to gather enough training data of animals taken by CCTV cameras to train an algorithm to recognise them accurately. We do gather data from customers (with permission) but less than 1% of these images contain animals of any kind, and most of these are bugs moving/flying close to the camera.

    - Normally, cameras are positioned and configured optimally to capture things like humans and vehicles. Most animals are far smaller, and so after the amount of cropping that has to be done to extract the small moving animal, the amount of detail there is often extremely poor.

    - Animals inevitably look a bit like humans in some ways, after all they have may of the same features (eye, limbs, ears etc.) and humans will inevitably be mistaken for animals. We want to avoid the situation where people are offended by our algorithm mistakingly classifying them as animals. To some extent this can't be avoided, but what we don't want to do is put out an inaccurate algorithm where this were a significant problem.

    The best way to capture animals at the moment is simply to turn off the human/vehicle detectors that exist in SecuritySpy, and instead use normal motion detection. This way, all motion, including animals, will be captured.
  • Yes, difficult, but not impossible and these are barriers that can be managed.
    For example, start with dogs and add a model per month of most common animals (bear, deer, cat, rodentia, etc). Make it a beta feature or opt in, and let the models improve.

    I'll echo that coming from Blue Iris with DeepStack API running in background was great at this and feels like a missing feature to another-wise very comprehensive and polished product.

    TLDR: Ben is right about difficulties, but I bet he can manage/resolve them!
  • Thanks for the feedback @jimmyjohnson and the vote of confidence :) The approach you outline sounds sensible to me.
  • Agreed that a more flexible AI for detection is really going to be mandatory for future competitive environments against other options. A halfway-decent detection tool would remove a huge number of false positives (which are most of my recording events) especially if it was customizable in some way. "person" and "car" aren't quite working for me. The deepstack stuff looks very interesting and seems to be "compatible" with inclusion as either a plug-in or sidecar to SS.

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!