Traditional 3D reconstruction pipelines—such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS)—rely on a fundamental assumption: the input images must be perfectly sharp. To reconstruct a scene, algorithms must establish multi-view consistency, matching features across different camera viewpoints.
However, real-world imaging sensors do not capture instantaneous snapshots. Instead, they open a physical shutter for a finite exposure window $\Delta t$, acting as a temporal integrator. If the camera shakes or an object moves rapidly during this window, high-frequency spatial details are smeared across pixels, resulting in motion blur.
When standard optimization techniques ingest these blurred frames, they experience a severe mathematical mismatch. As noted by Lin Zhu et al. in E2NeRF, without explicit handling, the optimization process treats the transient blur as a permanent, static feature of the 3D geometry. This causes the system to bake the artifacts directly into the scene representation, leading to cloudy "floater" artifacts, noisy geometry, and a severe loss of texture clarity (Peng Dai et al. in Hybrid Neural Rendering for Large-Scale Scenes with Motion Blur).
To solve this, modern research focuses on a powerful shift in perspective: motion blur is not destructive noise, but rather an integrated temporal record of the scene's motion. By explicitly modeling how blur is formed, or by augmenting frames with high-temporal-resolution sensors, we can invert this integration process and recover pristine 3D scenes.
Inverting the Blur via Forward Physical Modeling
The most direct way to handle motion blur is to modify the forward rendering pipeline to mimic the physical camera sensor. Instead of trying to pre-deblur the input images using 2D computer vision networks—which is an ill-posed problem that often introduces hallucinations—we can blur the renderer itself.
A blurry pixel is mathematically represented as the integral of radiance observed along a camera's continuous trajectory during the exposure time:
$$I_{\text{blur}} = \frac{1}{\Delta t} \int_{t_0}^{t_0 + \Delta t} \mathcal{R}(x, y, \mathbf{T}(t)) \, dt$$
where $\mathcal{R}$ is the rendering function and $\mathbf{T}(t)$ is the continuous camera pose trajectory.
[Continuous Trajectory T(t)] ──> Sample Virtual Poses ──> Render Sharp Views ──> Average (Integrate) ──> Simulated Blurry Frame
│
Compute Loss vs.
Real Blurry Image
By parameterizing the camera's motion as a continuous trajectory (such as a cubic B-spline), the renderer can sample multiple virtual camera poses along this path for a single exposure window.
In BAD-Gaussians, Lingzhe Zhao et al. apply this concept to 3D Gaussian Splatting. They jointly optimize the 3D scene parameters and the continuous camera motion trajectories. During the forward pass, the system renders several sharp images along the estimated trajectory and averages them to synthesize a simulated blurry frame. The loss function then compares this simulated blur directly with the real blurry input image.
Because the rendering pipeline explicitly accounts for the averaging process, the backpropagation gradient forces the underlying 3D representation to remain perfectly sharp. The only way the averaged virtual views can match the blurred input is if the underlying 3D geometry and textures are correct and clear. Similarly, Peng Wang et al. demonstrate in MBA-SLAM that embedding this continuous motion modeling into a dense visual SLAM tracking and mapping loop eliminates physical inconsistencies and ghosting artifacts.
Unlocking Time with Neuromorphic Sensors
While forward modeling works well for smooth, predictable camera trajectories, it struggles with highly complex, high-frequency jitter or abrupt movements where the exact motion path during exposure is ambiguous. To resolve this ambiguity, researchers leverage hybrid sensor setups that pair standard frame-based cameras with bio-inspired neuromorphic sensors, such as event or spike cameras.
Event cameras do not capture frames at fixed intervals. Instead, each pixel operates independently and asynchronously, emitting a digital "event" the moment it detects a change in log-intensity. This occurs at microsecond-level temporal resolution.
Standard Frame: [========= Integrated Exposure Window (Blurry) =========]
Event Stream: |..:...::...:...:..::..:...:..:...:..:...::..:...:..:...| (Microsecond changes)
When a standard camera suffers from motion blur, the accompanying event camera records the exact trajectory of brightness changes during that exact exposure window. The event stream provides the missing temporal derivative of the scene's appearance.
In Ev-DeblurNeRF, Davide Scaramuzza et al. combine a motion-aware NeRF with a learnable event response function. The events act as a mathematical bridge between the blurred frame and the latent sharp states. By integrating the events over time, the system can "un-integrate" the blurry frame, mapping the blurred pixel values back to their precise, sharp origins at any sub-frame timestamp.
This principle is equally effective when applied to alternative neuromorphic architectures. Yijia Guo et al. introduce SpikeGS, which pairs bio-inspired spike cameras with 3D Gaussian Splatting. Unlike event cameras, which measure relative changes, spike cameras continuously count incoming photons and emit high-frequency spike trains when a threshold is reached. This ultra-high temporal resolution completely bypasses the exposure-window limitations of traditional sensors, allowing high-fidelity 3D reconstructions from fast-moving platforms like drones or high-speed vehicles where standard RGB sensors fail completely.
Decoupling Non-Rigid and Object-Level Blur
The problem becomes significantly more complex when the motion blur is caused not by camera movement, but by dynamic, non-rigid objects moving within the scene—such as a waving hand or a moving face. In these scenarios, a single global camera trajectory cannot resolve the blur, because different parts of the scene are blurring in different directions simultaneously.
To solve this, frameworks must decouple the motion components. In BARD-GS, Yiren Lu et al. explicitly separate motion blur into two distinct components: global camera movement and independent local object deformation. By optimizing separate deformation fields for the underlying 3D Gaussians over time, the system can accurately map how an object changed during the exposure window, isolating the spatial smear from the true 3D structure.
When explicit tracking data is sparse, structural domain priors can fill in the missing high-frequency details. For instance, in Face Video Deblurring Using 3D Facial Priors, Wenqi Ren et al. demonstrate that 3D facial geometry and expressions can be reliably estimated even from heavily blurred frames. This estimated 3D structure serves as a stable spatial blueprint, providing a geometric anchor that guides the optimization process back toward a sharp, physically accurate reconstruction.