Event-Guided Dynamic Reconstruction under Extreme Motion Blur

@ The 1st Workshop on Physical World AI, NeurIPS 2026

Motion blur destroys the feature correspondences that reconstruction depends on, and Structure-from-Motion collapses with it. We use the microsecond timing of an event camera to recover them, and model the camera path as one continuous spline rather than a separate pose per frame.

Arpitsinh Vaghela¹, Aayush Atul Verma¹, Kaustav Chanda¹, Chi-Yao Huang¹, Duo Lu², Bharatesh Chakravarthi¹, Yezhou Yang¹

¹Arizona State University ²Rider University

Read the paper arXiv Code

Novel views under extreme blur. Ours stays sharp where the baselines smear or tear.

Method

Events restore the edges, a spline restores the motion

COLMAP finds keypoints as regions of strong intensity change, and extreme blur averages those edges away. On our benchmark it recovers no camera poses at all at severe or extreme blur, leaving every blur-aware method without the initialization it assumes. The Event-based Double Integral turns asynchronous events plus a blurred frame back into sharp intensity images, which replace the raw frames during SfM. Features reappear, and poses become estimable at a higher frequency than the blurred frames themselves.

The trajectory is then a single cubic B-spline in SE(3), not a pose per frame. Each control point supports a window of time and adjacent poses share them, so temporal coherence is a property of the representation rather than a penalty bolted onto the loss. That is where the accuracy comes from: pose error drops 35%.

Two 3D plots of the first thirty camera poses. The BAD Gaussians module groups poses into three tight clusters joined by long straight jumps, drifting from the dashed ground-truth curve. Our module places poses evenly along a smooth arc that tracks the ground truth. — First 30 poses on **sriracha-tree**. Per-frame refinement (top) clumps and jumps; our continuous spline (bottom) tracks the ground truth.

Three-stage framework diagram. Stage 1 fits a B-spline camera trajectory from control points. Stage 2 samples poses along the spline, renders and averages them into a synthetic blurred image, and applies blur and intensity losses through a dynamic mask. Stage 3 adds a deformation field conditioned on time to reconstruct the moving regions. — **Stage 1** fits the spline to the event-guided pose estimates. **Stage 2** refines it against both blurred frames and event-recovered intensities, with a dynamic mask keeping object motion out of the trajectory. **Stage 3** freezes the trajectory and learns a deformation field for the moving regions.

Results

Eight scenes, seven methods, extreme blur throughout

Deblurring PSNR

21.3121.83 dB

+0.52 dB

best baseline: EDyGS

Novel view synthesis, SI-PSNR

19.4020.42 dB

+1.02 dB

best baseline: BAD-4DGau. (w/ EV)

Absolute pose error

1.831.19 cm

35% lower

best baseline: EDyGS

Left to right: blurred input, 4DGaussians, 4DGaussians + MPRNet, BAD-4DGaussians, EDyGS, ours, sharp reference. All methods use the camera poses from our event-guided pipeline, because none of them can be run without them.

Method	Deblurring			Novel view synthesis
Method	PSNR ↑	SSIM ↑	LPIPS ↓	SI-PSNR ↑	SI-SSIM ↑	SI-LPIPS ↓
ED3DGS	19.59	0.5244	0.5365	18.13	0.5071	0.5521
4DGaussians	19.85	0.5321	0.5398	18.68	0.5129	0.5536
4DGaussians + MPRNet	19.66	0.5258	0.5344	18.69	0.5120	0.5442
DyBluRF	17.42	0.2717	0.6939	16.54	0.2597	0.7022
BAD-4DGaussians	18.65	0.4592	0.4866	17.88	0.4498	0.4961
4DGaussians (w/ EV)	20.35	0.5467	0.5199	19.07	0.5241	0.5322
BAD-4DGaussians (w/ EV)	20.98	0.5488	0.4358	19.40	0.5170	0.4556
EDyGS	21.31	0.5881	0.5058	19.07	0.5360	0.5227
Ours	21.83	0.5755	0.4091	20.42	0.5503	0.4244

best second third Averaged over eight Dycheck scenes at extreme blur. SI-* metrics allow up to 10% translation before scoring.

Paper

Cite this work

Download the PDF arXiv Code

@inproceedings{vaghela2026eventguided,
  title     = {Event-Guided Dynamic Reconstruction under Extreme Motion Blur},
  author    = {Vaghela, Arpitsinh and Verma, Aayush Atul and Chanda, Kaustav
               and Huang, Chi-Yao and Lu, Duo and Chakravarthi, Bharatesh
               and Yang, Yezhou},
  booktitle = {NeurIPS 2026 Workshop on Physical World AI},
  year      = {2026}
}