Recovering Temporal Dynamics from Single Blurry Images

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2118 papers160 conceptsfrozen 2026-10-04

When a camera shutter stays open while an object moves, the sensor performs a mathematical operation: it integrates light intensity over time. This process collapses a continuous three-dimensional spatio-temporal volume into a single two-dimensional image. In the computer vision community, this is known as motion blur.

Traditionally, motion blur was treated as a nuisance—a degradation to be removed. However, a growing body of research treats blur not as a loss of information, but as a compressed record of a high-speed event. The challenge lies in "unrolling" this integration to recover the sharp video sequence hidden within the smear.

The Problem of Temporal Ambiguity

Reconstructing a sharp video from a single blurred frame is a fundamentally ill-posed inverse problem. As Meiguang Jin, Givi Meishvili, and Paolo Favaro describe in Learning to Extract a Video Sequence From a Single Motion-Blurred Image, a single blurred image can correspond to an infinite number of potential sharp image sequences.

The most significant hurdle is directional ambiguity. A motion streak looks identical whether an object moved from left-to-right or right-to-left. Because the integration process is symmetric, the "arrow of time" is lost. Chaoyue Wang and Stephen Maybank note in Self-supervised Exposure Trajectory Recovery for Dynamic Blur Estimation that without additional temporal priors, it is impossible to distinguish the chronological order of the trajectory.

Unrolling the Blur into Video

To resolve this ambiguity, researchers have moved beyond simple deblurring toward photosequencing. Instead of producing one sharp image, these models "unroll" the blur into a sequence of frames.

Denys Rozumnyi and colleagues have pioneered methods like DeFMO (Deblurring and Shape Recovery of Fast Moving Objects) and Motion-from-Blur (Motion-from-Blur: 3D Shape and Motion Estimation of Motion-blurred Objects in Videos). Their approach involves disentangling the moving object from the background and using physical image formation models to estimate the object's 3D trajectory. To handle the directional ambiguity, they often employ "order-invariant" loss functions that evaluate multiple possible paths and optimize for the best fit, preventing the model from being penalized for a valid but time-reversed prediction.

Event Cameras as Temporal Bridges

A major breakthrough in this field has been the integration of event cameras (or neuromorphic sensors). Unlike standard cameras that capture frames at fixed intervals, event cameras record pixel-level intensity changes at microsecond resolution.

In Learning Event-Based Motion Deblurring, Zhe Jiang and Jimmy Ren demonstrate that these "spikes" of data provide the missing temporal link. Because events record the exact moment a change occurs, they act as a roadmap for the reconstruction. This allows models like Ev-DeblurVSR by Dachun Kai and Yueyi Zhang to achieve state-of-the-art restoration quality while running significantly faster than frame-only methods.

Chen Song and Chandrajit Bajaj further refined this in DeblurSR (DeblurSR: Event-Based Motion Deblurring Under the Spiking Representation) by using a "Spiking Representation" that mimics biological neurons. This approach captures sharp edges more effectively than traditional polynomial-based methods by treating pixel intensity as a series of discrete temporal events.

From Discrete Frames to Continuous Functions

The most recent evolution in this area is the shift from discrete frame sequences to Implicit Video Functions (IVF). Rather than predicting a fixed set of frames (e.g., 7 or 15), these models represent the video as a continuous parametric function of time.

In Recovering Continuous Scene Dynamics from A Single Blurry Image with Events, Xiang Zhang and Lei Yu introduced DeMo-IVF. By querying the function at any arbitrary timestamp within the exposure period, the system can reconstruct sharp latent images with unprecedented temporal flexibility. This effectively transforms a single blurry photograph into a high-frame-rate, continuous-time video, resolving the "double trouble" of low spatial resolution and motion blur simultaneously.

Papers in this version

Showing 60 of 118 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 78 papers
  2. Version 1 · 2026-10-04 · 118 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_direction_ambiguity_09f9a244,
  title = {temporal direction ambiguity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-direction-ambiguity/rdr_09f9a24418a74516ad0e4f2428a930db},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_recovering_temporal_dynamics_from_single_09f9a244_v1,
  title = {Recovering Temporal Dynamics from Single Blurry Images},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/recovering-temporal-dynamics-from-single-blurry-images/rdr_09f9a24418a74516ad0e4f2428a930db/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}