Pick a scene from the Physical AI Smart Spaces dataset on Hugging Face. Every camera is tracked together, giving one ID per object.
| Select | Scene | Split | Cameras | Type | Result | Video |
|---|
Expected: Camera_XXXX.mp4 files (1080p, 30 fps, same length) and a calibration.json with one sensor per video. map.png and ground_truth.json are optional.
Check where the cameras can see before running. Select a camera on the map or in the wall to show its view; switch cameras off to leave them out of the run.
Estimate from the loop-3 measurement of 1.3 frame-sets per second for one scene on an A100, at 6.77 Colab units per hour.
Every box is one 3D object, projected into each camera that can see it. Pick an object in any camera, on the floor plan or in the timeline to follow it everywhere.
Mockup data: these are real outputs of candidate D on this scene at frame 600 (0:20). Playback, the scrubber and exports are not wired up yet.
ground-truth lifetime predicted track following it (shade flips when the ID changes) ID change predicted, no ground-truth match
Matched by nearest centre on the floor within 1 m (1.5 m for vehicles), same class, every 3rd frame, ignoring blips under 0.4 s. An approximation for reading the timeline; the score panel uses the official evaluator.
Pick a box, a dot or a timeline bar.
Candidate D on the full scene, official 3D HOTA evaluator. A real run would show DetA, AssA and per-class rows here.
track1.txt uses the AI City Track 1 columns: scene, class, object, frame, x, y, z, width, length, height, yaw.