S2E: Spatio-temporal filtering of spike streams for motion-selective event generation

Spike cameras generate asynchronous, fovea-like spike streams that encode scene luminance with micro-second precision, while event cameras sense only motion-induced intensity changes. The complementary strengths of these two neuromorphic modalities motivate spikes-to-events (S2E) conversion, enabling a single spike sensor to emulate both foveal texture and peripheral motion pathways. Existing S2E methods, however, amplify reconstruction noise and miss slow variations because they produce events by frame-wise differencing of noisy, spike-derived intensities. We propose a fully integrated S2E pipeline that first denoises the spike stream with a double-stage intensity estimation with a Markov-random-field-based spatio-temporal filter and then applies spike-driven log-differencing to synthesise high-quality events at the native microsecond resolution. Experiments on three public data sets-PKU-Vidar-DVS, motVidarReal2020 and momVidarReal2021-demonstrate that the resulting event streams exhibit clearer motion contours and substantially lower background artefacts than those produced by raw event cameras or previous S2E approaches. The proposed framework thus offers a practical route toward single-sensor neuromorphic vision systems, unifying foveal intensity reconstruction and peripheral motion detection within one spike-camera platform.