ESAR: New math extracts clean images from event camera noise

Event cameras capture asynchronous polarity events when log-radiance changes exceed a threshold, producing signed temporal contrast measurements rather than full frames. The ESAR paper reformulates monocular event-based imaging as a synthetic-aperture inverse problem. Rather than reconstructing a latent pixel-time volume v (size N_p × N_t), the authors impose a geometric relation v = Pθ, where θ is a static ground-domain log-radiance field (size N_g). Aggregating events over finite intervals gives the linearized model APθ = b + η, where A is a temporal differencing operator, b holds signed binned event counts, and η represents errors. This exposes a synthetic-aperture structure: under near-nadir motion, successive projections become approximately shifted views of a common scene, but AP is ill-conditioned because it combines spatial averaging with temporal differencing.

To recover θ, the team uses regularized inversion. Numerical experiments on simulated data and real Falcon Neuro event data demonstrate that the θ-based formulation recovers coherent large-scale spatial structure compared to dynamic latent-image and learned event-reconstruction baselines, while suppressing fine-scale texture. This approach effectively turns noisy, asynchronous event streams into stable spatial representations, potentially enabling better imaging from fast-moving platforms like drones or satellites where traditional cameras struggle with motion blur.