Experiential Features

Experiential Features

These are the "new way to see" features: single-camera, mostly frontend, built on the same detection and motion model as the overlays.

Follow-cam

Lock onto a subject and a digital PTZ keeps them centred as they move. Press F on a selected subject, or ask the operator to follow. It rides the same motion model as the boxes and re-acquires the target through brief losses; it drops out cleanly when the subject is truly gone (a lost-grace timer), so it does not chase ghosts. On a tracking card the FOLLOW button replaces the old SIMILAR button.

Occlusion x-ray

See where a tracked subject is even while they are behind cover. When a subject passes behind an obstacle, Overseer holds and extrapolates it along its last velocity, drawn as a distinct predicted ghost, so you keep a read on it through the occlusion. Toggle it from the left rail (Vision) or the xray operator action.

Live narration

The active camera narrates itself aloud, as a subtitle: watch it or just listen. With one key (N) it describes the scene from the structured detection data: subjects by clothing colour, height and intent (loitering / hurrying / waiting), who is facing the camera, groupings, and people or vehicles entering and leaving. It rotates focus so it varies each time, and works with no vision model configured; when a vision model IS set, its visual one-liner is mixed in every so often. Toggle it from the left rail (Audio) or the narrate operator action.

Live enhance

Draw a box on any region for a crisp, photographic close-up. Multi-pass learned reconstruction fitted to the loupe, so detail is rebuilt rather than block-upscaled: a sharp optical zoom, not a cartoon. Press E, or ask the operator to enhance. Enhance closes on Esc and when you switch cameras. See Identity, Re-ID & Forensics.

Tactical god-view radar

A top-down mini-map that depth-locks each contact to its ground position (using Depth Anything), giving an at-a-glance read of the whole scene from above, including who is near whom.

Predictive ghosts

A translucent overlay of where each tracked subject is likely to be a few seconds ahead, with path-convergence detection that flags two paths about to meet before they do. Backed by the foresight engine (see Spatial & 3D).

Object-of-interest (OOI) tracking

Point at any object the detector does not auto-track (a bag, a box, a specific item) and Overseer will track that specific object visually as it moves, surfaced over the WebSocket as ooi updates.

HoloReel (3D video capture)

Record a few seconds of the scene in 3D and replay it like a video you can fly through. In the 3D Spatial view press HOLOREEL: the backend grabs a burst of genuinely distinct frames spread across about three seconds of real time (spatial.reel_span_s, default 3.0s), so the reel shows real motion instead of a sub-second window that looks frozen. Capture and reconstruction are interleaved (the depth inference between grabs plus a minimum per-frame stride does the spreading), and every frame is reconstructed and precomputed up front. A CAPTURING indicator shows while the frames are grabbed, then a BUILDING counter while the meshes precompute. Playback is just a visibility switch between finished meshes, so it flows like a video with no per-frame freeze, driven by a bottom scrub bar (play / pause and a slider). You can orbit and fly through the scene the whole time, pausing on any moment. If the live capture ring is not advancing for the source, HoloReel falls back to reading a consecutive burst straight from the source (sub-sampled across the window), so it never freezes on a single frame. It is the closest thing to holding a short volumetric clip of what happened.

The 3D view also has a plain LIVE mode (the reconstruction keeps refreshing from the live feed) and RECAPTURE for a single fresh build.

HoloReel replaces the earlier "Chronoscape" 2D-trail replay, which was hard to read on noisy monocular depth.

Social x-ray

See the invisible social layer of a scene on a single camera. Toggle it in the modules rail (VISUAL group) or press S.

• Each person gets an attention cone in the direction they are facing. Facing is derived from the pose (the shoulder line, the nose lean, and whether the eyes are visible), so it works even when someone is standing still, where a motion-only heading gives nothing. It falls back to motion heading when a pose is not available.
• People who are interacting are linked, but only with real evidence, so it stays conservative: ENGAGED (both facing each other, close), WATCHING (one facing the other, close), APPROACHING (closing in while moving). Mere proximity is not enough, so people standing side by side in a queue (facing the same way, not each other) are not linked. It is the moment-to-moment complement to the roster's long-term relationship graph.
• It is built to be readable, not twitchy: positions and links are eased and a link, once shown, lingers for about three seconds as it slowly fades, so you have time to follow it. The attention cone starts at eye level. The connection labels are light and spaced apart so the text on the lines never reads as bold or overlaps. Click a person and it shows ONLY that person's cone and links, nothing else, so a busy scene stays legible.
• Drive it from the operator: "social x-ray" (operator action social_xray).
• Prompt + light on the main thread: the pose model is warmed up off-thread when a camera connects, so cones and links appear within a pose cycle instead of after a multi-second first load. The overlay's per-frame pass is throttled to about 30fps (and uses O(1) track lookup), so it never competes with the live video even with several people on screen.

---

All of these are toggles or one-key actions, and are also reachable through the AI Operator (narrate, follow, xray, enhance, social_xray, walkthrough, track_object). Overlay toggles live in the left modules rail (Vision / Audio / Visual). POV shortcuts: N narration, F follow, E enhance, S social x-ray. Occlusion x-ray defaults off and remembers being turned off.