Quadratic Motion Modeling for Non-Linear Trajectories

39 concepts24 papers2 proposals

When we watch a video or track an object, our brains intuitively understand that things speed up, slow down, and move along curves. Yet, for decades, computer vision algorithms have relied on a simplifying assumption: that between any two frames, pixels move in a straight line at a constant speed. This is a first-order linear assumption.

While mathematically convenient, it fails spectacularly when capturing real-world dynamics. When an object accelerates or changes direction, linear models produce severe visual artifacts like ghosting, ringing, and double edges. To bridge this gap, a powerful paradigm has emerged: quadratic motion modeling. By moving from first-order linear approximations to second-order polynomial representations, algorithms can explicitly model acceleration. This shift allows systems to accurately trace curvilinear paths and changing velocities, unlocking dramatic improvements in video frame interpolation, deblurring, compression, and tracking.

The Intuition: Why Linear Models Fail

Consider a ball being thrown across a screen. If we only look at frame $t$ and frame $t+1$, a linear model draws a straight vector between the two positions. If we try to synthesize an intermediate frame at $t+0.5$, a linear interpolation places the ball exactly halfway along that straight line.

However, physics dictates that the ball moves along a parabolic arc, and its speed changes continuously due to gravity. A linear interpolation misses the curve entirely, placing the ball in an incorrect position. In video synthesis, this mismatch manifests as blur or a "ghosted" double image.

A quadratic motion model resolves this by incorporating a second-order term from a Taylor expansion of the trajectory. Instead of modeling displacement simply as a function of velocity ($\Delta x = v \cdot t$), it models it as:

$$\Delta x = v \cdot t + \frac{1}{2}a \cdot t^2$$

where $a$ represents the acceleration. This simple addition transforms a rigid straight line into a flexible curve that can bend and scale its velocity over time. As shown by Xiangyu Xu et al. in their work on Quadratic Video Interpolation, accounting for this acceleration term drastically reduces ghosting artifacts in complex scenes.

Linear Path (Constant Velocity):     [Frame 1] ----------> [Interpolated] ----------> [Frame 2]
Quadratic Path (With Acceleration):  [Frame 1] __        
                                                 \___      [Interpolated]
                                                     \____---------> [Frame 2]

Estimating Acceleration

The primary challenge of quadratic modeling is estimating the acceleration term $a$. A single velocity vector requires two frames; computing acceleration requires more information. Researchers have developed two main strategies to solve this:

1. Multi-Frame Temporal Windows

The most direct way to calculate acceleration is to look across multiple consecutive frames. By analyzing four consecutive frames instead of two, models can compute the change in velocity over time. This approach is used by Saikat Dutta et al. for space-time super-resolution, allowing the system to map complex trajectories.

To handle highly irregular kinematics without the extreme computational cost of higher-order polynomials, Meiqin Liu 0002 et al. proposed a piecewise approach in JNMR: Joint Non-linear Motion Regression for Video Frame Interpolation, concatenating multiple simpler quadratic stages over successive time intervals.

2. Exploiting Intra-Frame Information

What if we only have two frames, or even just a single image? Remarkably, acceleration can still be recovered. Motion blur itself contains a continuous record of an object's path during the camera's exposure window. Youjian Zhang et al. demonstrated in Exposure Trajectory Recovery From Motion Blur that pixel-wise quadratic polynomial constraints can unlock high-frame-rate video from a single blurry shot.

Similarly, Ugur Çogalan et al. combined standard inter-frame correspondences with intra-frame flow derived from motion blur in dual-exposure sensors to fit a quadratic model using only two consecutive capture events. Alternatively, Kui Jiang et al. developed an implicit framework (IQ-VFI) that uses a teacher-student distillation strategy to predict latent acceleration from just two sharp input frames.

Applications Across the Visual Spectrum

The benefits of modeling second-order motion extend across many domains in computer vision and video processing:

  • Video Deblurring: Real-world camera shake or object movement rarely follows a straight line during an exposure window. Wenqi Ren et al. showed that modeling pixel trajectories as quadratic functions rather than linear paths significantly reduces ringing artifacts, leading to much sharper video restoration in dynamic environments.
  • Rolling Shutter Correction: CMOS sensors capture images row-by-row. When a camera shakes or moves quickly, this sequential readout causes geometric distortions (the "jello effect"). Delin Qu et al. and Bingbing Zhuang et al. demonstrated that modeling camera and pixel trajectories as quadratic curves with constant acceleration accurately maps the erratic, jerky movements of handheld devices, allowing for precise rectification and alignment.
  • Video Compression: Modern video codecs save space by predicting future frames based on past ones (inter prediction). Traditionally, they assume constant velocity. By extending this to a Uniformly Accelerated Motion Model (UAMM), Feng Wu 吴枫 et al. reduced bit-rate prediction errors in the Versatile Video Coding (VVC) standard. Similarly, Ren Yang et al. integrated quadratic motion modeling into bi-directional frame prediction for learned video compression, achieving frame synthesis that rivals optimized x265 codecs at zero bit-rate cost.
  • Generative Flow Models: Beyond traditional video, the concept of quadratic trajectories is transforming generative modeling. While standard Rectified Flow models connect noise and data distributions using straight lines, Dogyun Park et al. introduced Constant Acceleration Flow. By defining the path between noise and data via a kinematic equation with a learnable acceleration term, they achieved higher-quality single-step generation through non-linear trajectories.

The Sweet Spot of Complexity

Why stop at quadratic? Why not use cubic, quartic, or higher-order polynomials? As explored by Bülent Sankur in early work on video codecs, quadratic modeling strikes an optimal balance between representational power and coding efficiency.

While cubic models—such as the one explored by Zhixiang Chi et al.—can capture variable acceleration, they require significantly more parameters and data to fit accurately. Quadratic models provide just enough flexibility to break the rigid constraints of linearity—capturing rigid rotations, linear deformations of quadratic surfaces, and constant accelerations—without introducing overfitting risks or severe computational overhead. By treating acceleration as a fundamental physical property of the visual world, quadratic motion modeling provides a robust, mathematically elegant framework for understanding dynamic scenes.

Key Foundational Work

A quadratic motion-based object-oriented video codec (2000) — contributed 2 concept elements to this direction.
Dynamic Color Tracking Based on Probabilistic Data Association (2004) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 24 papers
  2. Version 1 · 2026-10-04 · 38 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_quadratic_motion_modeling_b4922f1d,
  title = {quadratic motion modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quadratic-motion-modeling/rdr_b4922f1d47854686b534d691a4b4b2bd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_quadratic_motion_modeling_b4922f1d_v2,
  title = {quadratic motion modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/quadratic-motion-modeling/rdr_b4922f1d47854686b534d691a4b4b2bd/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake