E-MASC: Motion-Aware Dual-Coordinate Confidence Scoring for Event Camera Denoising

Abstract

Event cameras asynchronously report brightness changes with high temporal resolution and wide dynamic range, but background activity (BA) noise corrupts the sparse structures used by downstream perception. Conventional training-free filters primarily measure local support in observed coordinates, where fast motion can displace valid events beyond a fixed neighborhood. We propose Event-based motion-aware spatiotemporal confidence (E-MASC), which combines causal temporal recurrence, contrast maximization (CMax)-guided motion support, and spatial connectivity in raw and motion-aligned coordinates. The three cues are combined by a predefined weighted sum to produce a continuous confidence score, and a single operating threshold converts the score into binary decisions. E-MASC achieves a mean ROC AUC of 0.9657 on 20 complete DND21 sequences and an event-pooled F1 of 0.9213 on 66.0 million ED24 events. With identical E2VID settings, E-MASC attains SSIM/LPIPS values of 0.9424/0.1621 on ED24 and 0.9180/0.1602 on DND21, outperforming raw events and the tested conventional filters. Ablations over patch-grid size and CMax iterations quantify the tradeoff between ROC AUC and runtime and identify motion estimation as the main computational cost of the current offline implementation.