Introduction
In computer vision, the task of recognizing an object from an image requires a deep understanding of how its appearance changes with the observer's viewpoint. For polyhedral objects like cubes, this is straightforward as their features—edges and vertices—are fixed. However, for curved objects, the visible "edges" (known as limbs or silhouettes) are viewpoint-dependent. As an observer moves, these contours slide across the object's surface, making the mapping of all possible views a complex geometric challenge.
This paper presents the first algorithm for computing the exact aspect graph of piecewise-smooth curved objects defined by rational parametric patches. The aspect graph is a topological map that partitions the viewing space into regions where the object's silhouette remains stable. Crossing a boundary between these regions triggers a "visual event," where the structure of the silhouette qualitatively changes.
Methodology
The authors bridge abstract singularity theory with practical computation through several key techniques:
- Rational Parametric Modeling: Objects are modeled using rational parametric patches, allowing surfaces to be described via polynomial ratios.
- Singularity Theory: The research utilizes a catalogue of visual events (based on the work of Rieger and Koenderink) to define how silhouettes evolve. These include local events (swallowtails, beak-to-beak, and lip transitions) and multilocal events (triple points and tangent crossings).
- Numerical Continuation: To avoid the computational bottlenecks of purely symbolic algebra, the authors use numerical continuation to find starting points and a prediction-correction scheme to trace visual event curves across the viewing sphere.
- Visibility Handling: Ray tracing is employed to determine which parts of the mathematically defined silhouette are actually visible, accounting for self-occlusion.
Key Results
The algorithm successfully maps the boundaries of stable views on the Gaussian sphere. It provides clear visual evidence of capturing the three primary local transitions that define changes in silhouette topology:
- Swallowtail: A smooth contour develops a "nick" that splits into a junction and two cusps.
- Beak-to-beak: Two separate contour segments meet at a point and then merge into a single structure.
- Lip: A tiny closed loop of contour appears "out of thin air" on a previously smooth surface area.
By identifying these exact boundaries, the method ensures that no topologically distinct view of the object is missed, a significant advancement over previous approximate sampling methods.
Discussion and Future Work
While the current implementation focuses on orthographic projection (viewing from infinity), it establishes a rigorous framework for curved object representation. The move from symbolic "cylindrical algebraic decomposition" to robust numerical curve-tracing represents a major shift toward practical application in high-dimensional spaces.
Future work aims to extend this algorithm to perspective projection, which will increase the complexity of the underlying polynomial systems. Additionally, the authors look toward integrating these exact aspect graphs into real-time recognition systems to guide the matching of image features to 3D models.
Summary updated