Witness4D: 4D Scene Representations That Carry Their Own Evidence

A self-driving car follows another car. A bus pulls in front, and for three seconds the car ahead is completely hidden. Today's 4D reconstruction pipelines for driving scenes still draw that car, with exactly the confidence they use for a surface the cameras are looking at right now: nothing in the output separates what was measured from what was inferred. Those same pipelines also cannot say which parts of a scene are moving at all, unless a human annotator or a movable-class detector was paid to say so. Both failures come due at the same moment, and that moment is occlusion.

Witness4D proposes a 4D representation that carries its own evidence. Our goal decides what is moving from physics the reconstruction already contains, a rendered-velocity statistic and a temporal free-space contradiction, with no motion annotation, no movable-class detector and no per-scene tuning.