Motion Induced Ghosting in Temporal Filtering

14 concepts11 papers1 proposal

When processing video sequences, medical imaging streams, or real-time graphics, we often encounter unwanted noise, artifacts, or high-frequency jitter. A natural solution is to look across time. If a pixel's true value remains constant, averaging its value across multiple consecutive frames acts as a powerful low-pass filter, dampening transient noise and stabilizing the signal.

However, this approach relies on a fundamental assumption: that the spatial grid remains static relative to the underlying scene content. When an object moves, a fixed coordinate $(x,y)$ no longer tracks the same physical entity from frame to frame. Applying standard temporal filters under these conditions forces a mathematical conflict between noise suppression and motion fidelity, resulting in severe visual artifacts.

The Mechanics of Ghosting and Smearing

What happens when a temporal filter blindly aggregates pixels at fixed coordinates while an object moves across them? As Sylvain Paris demonstrates in Edge-Preserving Smoothing and Mean-Shift Segmentation of Video Streams, causal temporal smoothing blends current pixel values with past values that belong to entirely different physical objects or backgrounds. This manifests visually as ghosting—faint, trailing shadows of an object's previous positions that linger in the current frame.

This artifact is a recurring challenge across diverse domains:

  • Computer Graphics: In real-time ray tracing, performance constraints necessitate low per-pixel sample counts, resulting in heavy visual noise. As Tuo Chen et al. explain in Overview of Real-Time Global Illumination Rendering Methods without Pre-Processing, temporal denoisers used to clean up this noise frequently introduce ghosting or smearing during fast motion because they struggle to accurately reproject lighting information across rapidly changing frames.
  • Computer Vision Maps: When smoothing pixel-level prediction maps (such as semantic segmentation or saliency), Kang Dang et al. note in Adaptive Exponential Smoothing for Online Filtering of Pixel Prediction Maps that simple temporal filters incorrectly average moving foreground objects with the background, corrupting the spatial boundaries of the predictions.
  • Environmental Regularization: Techniques designed to stabilize video distorted by atmospheric turbulence often assume a static background to dampen random oscillations. Sung Ha Kang et al. show in Video stabilization of atmospheric turbulence distortion that when a dynamic object enters the scene, this static assumption is violated, causing severe ghosting and artifact propagation around the moving object.

The Signal Persistence Requirement

Beyond ghosting, temporal filtering introduces a more destructive failure mode: the complete erasure of high-velocity or small-scale features. For a temporal filter (especially a non-linear one like a median filter) to preserve a feature, that feature must fulfill a signal persistence requirement—it must occupy a specific spatial coordinate for a majority of the frames within the filter's temporal window.

When an object moves faster than this window allows, it is treated as transient noise and systematically eliminated.

Temporal Window (Frames 1-5)
Frame 1: [Object] [Background] [Background] [Background] [Background]
Frame 2: [Background] [Object] [Background] [Background] [Background]
Frame 3: [Background] [Background] [Object] [Background] [Background]

Result of Median Filter at any single position: [Background] 
(The object is completely erased because it never persists in one place)

This trade-off is particularly critical in medical imaging. In echocardiography, Isabelle Herlin et al. observe in Features Extraction and Analysis Methods for Sequences of Ultrasound Images that temporal smoothing designed to track slow-moving cavities inadvertently blurs or obscures fast-moving anatomical structures like cardiac valves. Similarly, Jens Frahm et al. demonstrate in On the Temporal Fidelity of Nonlinear Inverse Reconstructions for Real- Time MRI – The Motion Challenge that temporal median filters impose a strict upper bound on detectable velocity, completely erasing small, rapidly moving features because they fail to remain in a single spatial location long enough.

A similar phenomenon occurs in video processing and cinematography:

Breaking the Static Constraint

To resolve the conflict between noise reduction and motion preservation, researchers have developed strategies to break the static grid constraint.

Spatiotemporally Adaptive Kernels

Instead of filtering along a straight temporal line at fixed coordinates, filters can adapt to the local structure of the video. Henrik Malm et al. utilize a 3D structure tensor in Adaptive enhancement and noise reduction in very low light-level video to construct adaptive smoothing kernels. These kernels bend and span across both space and time along the direction of least variance (the actual motion trajectory), allowing aggressive denoising in low-light video without inducing motion blur or edge degradation.

Dimension Decoupling and Alternative Spaces

Another strategy is to avoid temporal filtering entirely where motion is high, or to change the operational space. For example, Isabelle Herlin et al. handle high-velocity components by relying strictly on spatial-only gradients, reserving spatio-temporal smoothing exclusively for slow dynamics.

Alternatively, Yohann Tendero et al. address non-uniformity correction in infrared cameras—a problem traditionally solved by averaging frames over time during camera motion—in Efficient single image non-uniformity correction algorithm. The authors equalize histograms across adjacent columns within a single frame. By treating spatial columns like video frames, they eliminate the need for temporal integration entirely, completely bypassing motion-induced ghosting.

Key Foundational Work

Features Extraction and Analysis Methods for Sequences of Ultrasound Images (1992) — contributed 3 concept elements to this direction.
Adaptive enhancement and noise reduction in very low light-level video (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_motion_induced_ghosting_f59829b3,
  title = {motion-induced ghosting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-induced-ghosting/rdr_f59829b34d4a4c119ee79ceaadc8f135},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_motion_induced_ghosting_f59829b3_v1,
  title = {motion-induced ghosting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/motion-induced-ghosting/rdr_f59829b34d4a4c119ee79ceaadc8f135/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake