The best surveillance AI will never come from one company or one dataset. It will be built by a community.

Heimdall Vision is a community project building one shared, open object-detection model for CCTV and surveillance footage. Sign in to contribute your own camera images and labels, or just use the trained model — completely free, either way.

How it works

1

Install the model

Run the current trained model on your own camera setup.

2

Upload when it's wrong

See a miss or a low-confidence detection? Upload that image and contribute it.

3

Annotate

Label what the model got wrong or missed.

4

We retrain

The shared model is retrained on everyone's contributions at regular intervals.

Why this isn't just another annotation tool

Plenty of tools can help you label images and train a model. That's not the hard part. The hard part is that any single home, business, or camera setup only ever sees a narrow slice of the world — one country, one climate, a handful of camera angles. A model trained on that alone will confidently misread anything it hasn't seen before: a delivery drone mistaken for a bird, a cruise ship mistaken for a car, because "large novel thing" was never a training example. Thousands of independent contributors, each adding what their own cameras uniquely see, is how that gap actually closes — something no single person or company can do alone.

Your data

Your images are never browsable by other users — only you, and admins reviewing new contributors before they're trusted, can see them. By uploading, you grant Heimdall a license to store, process, and use that footage to train the one shared community model. Right now that's a single project focused on CCTV and surveillance footage, not a menu of separate options.

You're responsible for having a lawful basis to capture and share whatever you upload — including, where your local law requires it, giving notice to and obtaining consent from anyone identifiable in the footage (passersby, license plates, and similar). Don't upload footage you don't have the right to share.

You can delete your own images at any time. If you're pictured in someone else's upload and want it removed, or want to know what data we hold about you, contact us (below) and we'll act on it.

What actually helps

Not raw volume — diversity. A thousand more daytime images of the same suburban driveway teach the shared model very little it doesn't already know. Night footage, rain, snow, IR/thermal cameras, unusual angles, and cameras from parts of the world that are underrepresented today are exactly the conditions that make the community model meaningfully stronger for everyone, including you.

Get started today

Sign in, contribute footage from your own cameras, or just use the trained model — free either way.

Sign in to get started