MEP-3DGS: Multi-Scale Pyramid Optimization and Event-Aware Masking for Efficient Motion-Deblurred 3D Gaussian Splatting

Motion blur weakens multi-view geometric consistency, causing 3D Gaussian Splatting (3DGS) to produce inaccurate geometry, floating artifacts, and blurred structures due to missing high-frequency information. Event cameras provide microsecond-level temporal cues, motivating an investigation of how event supervision can be incorporated into coarse-to-fine Gaussian optimization while accounting for sensor noise. We propose MEP-3DGS, a multi-scale event-aware progressive framework for motion-deblurred 3D Gaussian reconstruction. A coarse-to-fine pyramid optimization strategy first establishes stable global geometry before recovering high-frequency details. A piecewise-linear event warm-up mechanism gradually introduces event constraints during Gaussian densification, while an event-activity mask guides spatially selective edge refinement to enhance motion boundaries and reduce unnecessary supervision in low-activity regions. On the E2NeRF benchmark, MEP-3DGS achieves peak signal-to-noise ratios (PSNRs) of 30.31 dB for deblurred-view reconstruction and 27.89 dB for novel-view synthesis, together with competitive structural and perceptual quality. Real-world results are evaluated using a no-reference image-quality metric and qualitative comparisons. Training takes approximately 30 min per scene, with rendering speeds above 250 frames per second and peak memory usage below 5 GB on an RTX 3090 GPU.