Continuous Trajectory Modeling for Dynamic Vision

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

Version 1 of 2198 papers264 conceptsfrozen 2026-10-04

Traditional computer vision treats video as a "flipbook"—a sequence of discrete snapshots captured at fixed intervals. While this works for static scenes, it fails to capture the physical reality of motion. Objects do not teleport between frames; they follow smooth, continuous paths.

A growing body of research is shifting from this discrete-time paradigm toward continuous-time trajectory modeling. By representing motion as a differentiable mathematical function—typically using parametric splines or differential equations—researchers are solving long-standing problems in motion blur, rolling shutter distortion, and real-time dynamic reconstruction.

The Parametric Curve as a Motion Prior

The core mechanism of this research direction is the replacement of independent per-frame poses with a global trajectory function $T(t)$. Instead of optimizing thousands of individual coordinates, systems optimize a small set of control points for a parametric curve.

Jongmin Park and Munchurl Kim demonstrate in SplineGS that Cubic Hermite Splines provide an ideal balance for dynamic 3D Gaussian Splatting. Unlike global polynomials, these splines offer local support: adjusting a control point at $t=10$ does not affect the trajectory at $t=1$. This locality is crucial for stability; it prevents the "wobble" often seen when trying to fit a single complex motion with a high-degree polynomial.

Other researchers, such as Jihwan Yoon, argue for the use of Non-Uniform Rational B-Splines (NURBS). In SplineGS: Learning Smooth Trajectories, Yoon shows that NURBS provide a stronger temporal inductive bias, ensuring that reconstructed motion remains physically plausible even when visual data is sparse or occluded.

Solving the Realities of CMOS Sensors

Most modern cameras use rolling shutters, which capture image rows sequentially rather than all at once. When the camera moves quickly, this creates "jello" distortion because the camera's pose changes between the first and last row of a single frame.

Moyang Li and Peidong Liu address this in USB-NeRF by modeling the camera path as a continuous B-spline in $SE(3)$ space. Because the trajectory is continuous, the system can query the exact camera pose for every individual row's timestamp. This approach, pioneered in early work by Luc Oth and Roland Siegwart in Rolling Shutter Camera Calibration, allows for sub-frame precision that discrete models simply cannot achieve.

Similarly, event cameras—which record asynchronous brightness changes at microsecond resolution—require continuous-time modeling to be useful. Elias Mueggler and Davide Scaramuzza showed that by treating the camera path as a continuous curve, one can integrate event data at its native temporal resolution, avoiding the need to group events into artificial, low-frequency "frames" (Continuous-Time Trajectory Estimation).

Turning Blur into Information

Motion blur is traditionally viewed as noise. However, in a continuous-time framework, blur is actually a rich temporal record. A blurred pixel is the mathematical integral of a trajectory over the camera's exposure time.

In Human from Blur, Yiming Zhao and Jie Song use this insight to track 3D human poses. By optimizing a continuous trajectory such that its "re-blurred" projection matches the input image, they can recover the specific arc of a limb within a single exposure. Dongwoo Lee applies a similar logic to radiance fields in ExBluRF, using Bézier curves to parameterize the 6-DOF camera path during exposure to reconstruct sharp 3D scenes from "messy" handheld video.

Efficiency and the End of the MLP Bottleneck

Before the rise of explicit splines, dynamic scenes were often modeled using implicit neural networks (MLPs) that took $(x, y, z, t)$ as input. While flexible, these were slow to train and render.

The shift toward explicit dynamic splines has led to massive performance gains. Junoh Lee and Hae-Gon Jeon argue in Fully Explicit Dynamic Gaussian Splatting that using mathematical interpolators for translation and rotation maintains the blistering rendering speeds of static 3DGS. By bypassing MLP queries, SplineGS achieves rendering speeds up to 8,000x faster than previous implicit methods.

From Pixels to Trajectory Fields

The ultimate evolution of this idea is the "Trajectory Field." Instead of estimating flow between two frames, researchers like Xinhang Liu and Yu-Wing Tai propose mapping every pixel in a video to a continuous 3D trajectory.

In Trace Anything, they use a feed-forward network to predict B-spline control points for every point in a scene. This turns a video into a queryable 4D volume where a pixel's position can be retrieved at any arbitrary timestamp—even those between the original frames. This continuous representation enables "emergent" capabilities like zero-shot motion forecasting and smooth robotic manipulation, proving that when we treat time as a continuum, the "magic" of vision doesn't have to slow down.

Papers in this version

Showing 60 of 198 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 246 papers
  2. Version 1 · 2026-10-04 · 198 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_time_trajectory_modeling_b3424d13,
  title = {continuous-time trajectory modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-time-trajectory-modeling/rdr_b3424d133d66481eade50527605704f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_continuous_trajectory_modeling_for_dynam_b3424d13_v1,
  title = {Continuous Trajectory Modeling for Dynamic Vision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-trajectory-modeling-for-dynamic-vision/rdr_b3424d133d66481eade50527605704f5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}