Event cameras don't take pictures. Each pixel fires asynchronously whenever the log-brightness it sees crosses a threshold, so you get a stream of signed contrast blips, not frames. This work treats reconstructing a scene from those blips as a synthetic-aperture inverse problem, the same broad idea radar uses to build an image from a moving sensor.

The trick is to solve for a fixed ground-level radiance field rather than a full pixel-by-time volume, tying the two together through a geometric projection. When the camera moves nearly straight down, consecutive views are roughly shifted copies of the same scene, though the combined operator mixes spatial averaging with temporal differencing and ends up badly conditioned, hence regularized inversion. On simulated data and real Falcon Neuro recordings, it recovers coherent large-scale structure while smearing fine texture.

That texture tradeoff is stated plainly in the abstract, and the underlying math deserves the paper itself.