Abstract
Detecting interest points in vector flow fields—such as vortices, sources, and sinks—is critical for analyzing fluid dynamics, meteorology, and computer vision. Unlike scalar images where SIFT-like methods are standard, vector fields require descriptors that account for directionality and background laminar flow. This paper introduces a framework that models local flow using a complex monomial basis. This basis provides mathematical invariance to rotation and scale, allowing for the robust detection of singular patterns even when embedded in strong background currents.
Introduction
In computer vision, identifying "interest points" like corners or blobs in scalar images is a well-established task. However, vector flow fields—representing the velocity and direction of fluids, air, or fingerprints—present unique challenges. Existing detection methods often rely on fixed-size templates or simple affine models that fail when patterns are rotated, scaled, or moving against a "laminar" (constant) background flow. This research bridges the gap by providing a multi-scale, rotation-invariant detector specifically designed for the directional nature of vector data.
Methodology: Complex Monomial Basis
The core of the proposed method is a higher-order model for flow fields based on complex analysis. Local 2D flow is represented as a complex-valued function $f(z)$, which is approximated using a linear combination of complex monomials ($z^k$ and $iz^k$).
These functions form an orthogonal basis, making them highly efficient for modeling smooth, natural motions. A key mathematical insight is that these basis functions are eigenfunctions of the rotation operator. Consequently, when a flow field rotates, its projection coefficients change in a predictable, linear way. This property allows the calculation of "Principal Orientations" for flow patterns, analogous to how SIFT identifies dominant gradient directions.
The Detection Pipeline
To isolate "singular points" where the flow forms a distinct pattern, the authors calculate "singular energy." By ignoring the $k=0$ coefficients—which represent constant background flow—the detector can isolate a vortex even if it is being "carried" by a larger current.
The algorithm follows a SIFT-like pipeline:
- Scale-Space Construction: Creating octaves of the flow field via down-sampling and varying the variance ($\sigma$) of a Gaussian weighting function.
- Energy Projection: Convolving the field with the basis functions to produce energy maps.
- Local Maxima Detection: Finding singular energy peaks across both spatial coordinates and scales.
Experimental Evaluation
The method was evaluated on diverse datasets:
- European FLUID Project: Sequences containing multi-scale vortices in fluid mechanics.
- SSEC Satellite Imagery: Velocity fields from weather systems used to detect cloud divergence and large-scale atmospheric vortices.
- JHU 3-D Turbulence Dataset: 2D slices from a massive turbulence database to test pattern clustering.
Key Results and Insights
The detector successfully identified vortices in turbulent flows and distinguished between cloud disappearance (sinks) and sudden divergence in satellite data. A major advantage is the separation of background flow; by mathematically removing the $k=0$ term, the method "sees through" linear currents that typically obscure singular patterns. Furthermore, because the basis functions are eigenfunctions of rotation, the authors can rotate the coefficients rather than the entire field, significantly reducing computational overhead.
Conclusion
This framework provides a robust, scale, and rotation-invariant approach to vector field analysis. While currently constrained to 2D fields, it offers a significant advancement over traditional template-matching techniques. Future work aims to extend this methodology to 3D flow fields, where singular structures become more complex three-dimensional entities.
Summary updated