Abstract
Subjective contours are illusory edges perceived by the human visual system in the absence of physical luminance gradients. This paper establishes a mathematical foundation for these contours by deriving their shape from a set of consistency axioms. By requiring that a contour-filling algorithm be extensible, independent of point replacement, and invariant under affine transformations (viewpoint independence), we prove that the unique consistent shape for these curves is a conic section, specifically a parabola in the case of linked point-tangents. This geometric approach provides a computational model that aligns with classic psychological phenomena like the Kanizsa triangle.
Introduction
In both biological and computer vision, the problem of "missing" data is pervasive. Edges are frequently occluded or faint, yet the human visual system seamlessly "fills in" these gaps with subjective contours. While this phenomenon has been studied extensively in psychology, previous computational models have often relied on ad-hoc heuristics or complex biological simulations of neural networks.
This research seeks a principled, mathematical answer to a fundamental question: given two points and their associated directions (tangents), what specific shape should a curve take to bridge the gap? We propose that the answer lies in the geometric constraints of the physical world rather than just biological implementation.
Methods
We employ an axiomatic approach to define the requirements for a "consistent" contour-fitting algorithm. Three core axioms are proposed:
- Extensibility: If a new data point is added that already lies on the predicted curve, the overall shape of the curve should remain unchanged.
- Point Replacement: An endpoint of a curve can be replaced by any other point further along that same curve without altering the curve's global geometry.
- Affine Commutativity (Viewpoint Independence): The algorithm must produce consistent results regardless of linear transformations such as rotation, scaling, or skewing. This ensures the model is robust to changes in the viewer's perspective relative to a planar object.
Using homogeneous coordinates and projective geometry, we demonstrate that any curve satisfying these axioms must be a conic section.
Results
The primary result of this derivation is that for two "linked" point-tangents—where the tangent lines intersect at a common point—the only consistent shapes are a parabola or a pair of straight lines.
As shown in Figure 1, the model successfully recreates the curved edges perceived in the Kanizsa triangle. While the triangle is often perceived with straight edges, slight variations in the inducer shapes can lead to curved subjective contours that our parabolic model accurately predicts.
For more complex scenarios where points are not directly linked, we propose a "double parabola" construction—two parabolas joined tangentially. This maintains the required consistency while allowing for more flexible interpolation.
Figure 4 illustrates the algorithm's ability to interpolate curves between various sets of synthetic point-tangent pairs, demonstrating the robustness of the conic section approach across different configurations.
Discussion and Limitations
This work suggests that any visual system—human or machine—that assumes occluding contours are planar and distant must produce parabolic shapes to remain mathematically consistent. However, several limitations remain. The current proof relies on parallel projection, which assumes the viewer is far from the object; it does not fully account for perspective effects (central projection) at close range. Additionally, the theory assumes planarity of the real-world contours. Future work will explore the extension of these axioms to multiple points and the derivation of curvature-continuous paths, likely involving a global fit of parabolic segments.
Summary updated