FAQ

Overseer is an AI computer-vision and spatial-intelligence platform. It turns ordinary camera feeds into structured understanding: real-time detection and tracking, monocular depth, 3D reconstruction, cross-camera re-identification and event analytics.

Security operators, researchers and builders who want on-device situational awareness and 3D scene understanding without sending footage to the cloud.

Yes, the project lives on GitHub under the MIT license. See the repository for the current source and issues.

No. All inference runs on the host and models are vendored locally, so Overseer can run fully offline / air-gapped. Footage, embeddings and events stay in a local SQLite store on disk.

Yes, with reduced throughput. Detection and tracking degrade gracefully to CPU; depth and super-resolution are much faster on an NVIDIA GPU (CUDA + FP16).

The service binds to localhost with no auth by default. For networked or multi-user setups, front it with a reverse proxy that enforces your auth (bearer token, mTLS, SSO).

RTSP and ONVIF network cameras (with LAN discovery), local video files and single images. Looped clips are handy for testing.

Monocular depth (Depth Anything V2) is relative and generalises well. It is not metric without calibration; a guided calibration flow on the roadmap upgrades it to true scale.

Yes. DETECTION class filters are gated at the detector, so turning off a class removes it from detection, tracking, Re-ID and analytics, real load shedding, and the setting persists.

Open an issue on the GitHub repository with steps to reproduce or a clear description of the request.

Docs are Markdown files under website/src/docs/. Add a file with front matter (title, order, tag: docs) and it appears in the sidebar automatically.