In computer vision, describing the "shape" of an object usually requires calculating its geometric properties, such as curvature, surface normals, or its topological skeleton. Traditionally, these properties are computed using differential geometry—calculating derivatives of the boundary or distance field. However, derivatives are notoriously sensitive to noise; a single misplaced pixel can cause a local slope to swing wildly, leading to "spiky" skeletons or unstable curvature estimates.
A recurring theme across these papers is the shift from differential measures to integral measures. By replacing local derivatives with integrals over a neighborhood or the entire shape, researchers have developed a framework for shape analysis that is inherently resistant to noise, discretization errors, and missing data.
The Smoothing Power of Integration
The core intuition behind this research direction is that integration acts as a natural low-pass filter. While a derivative amplifies high-frequency noise, an integral averages it out.
Jacques Lévy Véhel demonstrated this early on by using the Favard length—an integral measure of intersections—to segment noisy medical images where local derivatives failed. This philosophy is echoed in the work of Wei-Yang Lin, who replaced noise-sensitive differential operators in 3D face recognition with "summation invariants." By using weighted sums of coordinates within a window, the system achieves a high signal-to-noise ratio even with imperfect spatial data.
Integral Curvature: Beyond Pointwise Derivatives
Standard curvature is defined by the second derivative of a curve. If the curve is represented by a jagged string of pixels, the second derivative becomes chaotic.
To solve this, researchers like Hendrik J. Weideman and Amir Jamaludin utilize integral curvature. Instead of looking at the change in tangent direction at a single point, they place a small circular window over the point and measure the area of the shape that falls within that window.
- If the area is exactly half the circle, the boundary is a straight line.
- If the area is more than half, the boundary is convex.
- If the area is less than half, it is concave.
This area-based ratio provides a stable, smoothed signature of the contour. Weideman used this to identify individual dolphins from the "nicks and notches" on their fins, while Jamaludin applied it to measure spinal curvature in low-resolution medical scans. In both cases, the integral approach allowed for precise geometric measurement from data that was too noisy for traditional calculus.
Average Outward Flux for Stable Skeletons
The "skeleton" or medial axis of a shape is the set of points equidistant from the boundaries. Finding these points usually involves looking for singularities in a distance field—points where the gradient is undefined. In practice, boundary noise creates thousands of "fake" singularities, resulting in a messy, branched skeleton.
Morteza Rezanejad, Gregory Dudek, and Allen R. Tannenbaum address this by treating skeletonization as a physical "flux" problem. Instead of searching for local singularities, they calculate the Average Outward Flux (AOF) of the distance gradient field.
By integrating the flow of the gradient through a small neighborhood, they can identify the true topological backbone of a shape. As Tannenbaum explains, the flux remains strictly negative at skeletal points but vanishes elsewhere. This integral formulation allows robots to map environments and recognize objects even when their sensors produce jagged, incomplete boundaries Average Outward Flux Skeletons for Environment Mapping and Topology Matching.
Global Moments and Volumetric Invariants
When local features are too corrupted to trust, the research shifts toward global integral measures. These methods describe a shape by integrating properties over its entire volume or area.
Zsolt Sánta and Zoltan Kato demonstrated that 3D meshes can be registered (aligned) without finding point-to-point correspondences by simply optimizing for the alignment of global geometric moments. Similarly, Jacek Tabor showed that the "circularity" of a shape can be robustly measured using the Kullback-Leibler divergence of its global spatial distribution.
Because these measures depend on the aggregate sum of all pixels or voxels, a missing chunk of data or a noisy edge has a negligible impact on the final descriptor. This makes global integrals the tool of choice for matching deformed shapes or objects viewed from different, distorting perspectives Noise-Resistant Affine Skeletons of Planar Curves.