Calibration-driven modulation transfer function estimation for event-based cameras

Pixels in event-based cameras (EBCs) report log-intensity changes asynchronously and operate in high dynamic range with microsecond temporal resolution. However, event streams do not correspond to radiometric values, precluding the direct application of standard electro-optical metrics such as slanted-edge modulation transfer function (MTF). In this work, we characterize three per-pixel event statistics-positive event accumulation (PEA), time-to-first-event (TFE), and inter-event timing (IET)-computed from the output of an EBC observing an illumination stimulus temporally modulated by square waves with known intensity amplitudes to learn the mapping between the PEA/TFE/IET values and the intensity values. The intensity-equivalent image reconstructed from a temporally modulated slanted-edge target by inverting the "intensity-to-event statistics" curve replaces the conventional frame samples in the standard ESF/LSF/MTF pipeline. A key aspect of the approach is the use of per-pixel calibrations, where a controlled, dynamically modulated LED stimulus is used to map event statistics values at every pixel. Reversing these per-pixel mappings helps overcome fixed-pattern noise (FPN), which we experimentally show to be the dominant source of degradation. We experimentally verified that MTF curves reconstructed from event-based and frame-based camera data are in strong agreement.