Bayesian selection of scaling laws for motion modeling in images

Patrick Héas, Étienne Mémin, Dominique Heitz, Pablo D. MininniICCV 2009 · DOI

Introduction: The Challenge of Turbulent Optic-Flow

Estimating motion in image sequences, known as optic-flow, is a fundamental problem in computer vision. However, it is inherently ill-posed due to the aperture problem—having more unknowns than equations. Traditional solutions, such as the Horn-Schunck method, rely on generic smoothness priors. While effective for rigid body motion, these priors fail significantly when applied to turbulent fluids, such as atmospheric clouds in satellite imagery or smoke in laboratory settings. Turbulent motion is characterized by multi-scale structures and specific statistical patterns that do not conform to simple smoothness assumptions.

Complex turbulent motion field estimated from image sequences.

Furthermore, standard regularization techniques require manual tuning of a weight parameter ($\alpha$), a process that is often subjective and time-consuming. This paper introduces a framework that replaces arbitrary smoothness with fluid physics and automates parameter selection through Bayesian inference.

Methodology: Physics-Based Regularization and Bayesian Evidence

The authors propose a hierarchical approach to motion estimation that integrates Kolmogorov’s turbulence theory into the regularization process.

1. Physical Regularization via Scaling Laws

Instead of assuming uniform smoothness, the model constrains motion increments—the change in velocity over distance—to follow physical scaling laws. By using the second-order structure function as a quadratic constraint, the estimated motion is forced to behave as a self-similar physical process across different spatial scales.

2. Bayesian Model Selection

To eliminate manual parameter tuning, the scaling law parameters (slope $\zeta$ and factor $\beta$) are treated as variables to be inferred. The authors employ Bayesian evidence maximization (Type-II Maximum Likelihood). This method calculates a "log evidence" score that incorporates an Occam factor, which penalizes overly complex models to prevent over-fitting noise.

3. Optimization

The resulting constrained minimization problem is solved using Lagrangian duality and the Uzawa algorithm. This approach effectively decomposes the global problem into a series of first-order regularizers acting across multiple scales, ensuring computational tractability.

Results and Discussion: Accuracy and Statistical Fidelity

The method was validated using both synthetic 2D turbulence (generated via Direct Numerical Simulation) and real-world meteorological data from METEOSAT.

Performance on Synthetic Data

On synthetic datasets where ground truth is known, the proposed method achieved a Root Mean Square Error (RMSE) of 0.091. This significantly outperforms other fluid-dedicated regularizers, such as those by Corpetti et al. (0.134) and Yuan et al. (0.096).

Statistical Fidelity

A key strength of this approach is its ability to capture the physical power-law behavior of turbulent flows. As shown in the statistical analysis of motion increments, the method maintains high fidelity to the underlying physics, whereas standard methods tend to over-smooth small-scale details.

Normalized PDFs of motion increments demonstrating the capture of physical power-law behavior.

Real-World Application

When applied to METEOSAT satellite imagery of the North Atlantic, the method successfully estimated an energy flux ($\hat{\epsilon} \simeq 0.79 \times 10^{-5} m^2s^{-3}$) that aligns with independent aircraft-based physical measurements. This demonstrates the method's utility in bridging the gap between computer vision and quantitative fluid mechanics.

Conclusion

By treating the image as a physical sensor and leveraging Bayesian selection, this research provides a robust, parameter-free method for fluid motion estimation. While the current model assumes homogeneous and isotropic turbulence, future work aims to extend this framework to 3D volumetric data and more complex, non-isotropic flows.

Summary updated

BibTeX
@inproceedings{heas2009bayesian,
  author = {Patrick Héas and Étienne Mémin and Dominique Heitz and Pablo D. Mininni},
  title = {{Bayesian selection of scaling laws for motion modeling in images}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2009},
  doi = {10.1109/iccv.2009.5459353},
}
Figure