Event-Guided Dynamic Reconstruction under Extreme Motion Blur

Dynamic scene reconstruction relies on accurate feature correspondences across frames. Under motion blur, these correspondences become unreliable, causing SfM pipelines to fail. To address this challenge, we propose a framework for event-guided dynamic reconstruction under extreme motion blur that leverages the high temporal resolution of event camera to guide camera trajectory estimation and scene deblurring. By fusing asynchronous event data with blurry images, our method recovers sharp dynamic scenes from blurred images. To model smooth, continuous motion, we represent the camera trajectory as continuous B-splines, jointly optimized with dynamic 3D Gaussians. Experiments show that our method improves PSNR by 0.52 dB for deblurring and 1.02 dB for novel view synthesis under most challenging blur level in our evaluation. The proposed continuous trajectory formulation further reduces pose error by 35\% compared to discrete alternatives. By integrating event-based sensing with continuous trajectory modeling, our framework enables robust dynamic scene reconstruction under extreme motion blur.