Introduction
A fundamental challenge in artificial intelligence is bridging the gap between precise 3D geometric data and the "fuzzy" spatial terms used in natural language. While a computer perceives a scene as a collection of XYZ coordinates, humans use relative terms like "near," "above," or "between." This research addresses how to translate geometric scene descriptions into cognitively plausible spatial relations using fuzzy logic and geometric approximations.
Computational Framework
The proposed model operates on a multi-level hierarchy to balance computational efficiency with cognitive accuracy.
Object Idealization and Coordinate Systems
To mimic human perception, objects are not treated as complex meshes. Instead, the system uses a hierarchy of approximations, primarily the Bounding Right Parallelepiped (BRP). For projective relations such as "left" or "behind," the system must account for the intrinsic orientation of the Reference Object (REFO). This is achieved through a series of transformations that establish a local coordinate system centered on the REFO.
Fuzzy Evaluation via Cubic Splines
The core of the model is the transition from binary logic to a continuous "degree of applicability." Instead of a hard cutoff for whether an object is "near" another, the system employs cubic spline functions. These functions map physical distance and angular deviation to a value between 0 and 1, where 1 represents perfect applicability and 0 represents no applicability.
Key Insights
The research highlights several critical aspects of spatial cognition:
- Relative Scaling: The definition of "near" is not absolute; it scales based on the dimensions of the reference object. This allows the model to correctly interpret "near the building" differently than "near the fire hydrant."
- Topological vs. Projective Relations: The model distinguishes between topological relations (e.g., at, near), which are primarily distance-dependent, and projective relations (e.g., above, front), which require both distance and directional calculations.
- The "Between" Relation: This is treated as a special case involving two reference objects. The degree of applicability is calculated as the arithmetic mean of being "in front of" the first object (from the perspective of the second) and vice versa.
Results and Implementation
The model was integrated into the VITRA (Visual TRAnslator) workbench, a system designed to generate natural language descriptions from visual scenes. The workbench allowed for the visualization of "applicability regions"—clouds of space where a specific relation holds true to varying degrees. This validated that the highest degree of applicability for a term like "above" is directly over the center of an object, tapering off as the angle increases.
Conclusion and Future Directions
While the model successfully implemented a functional 3D spatial reasoning system, several areas for improvement remain. The spline functions were manually tuned and require broader psychological validation to match human intuition across different contexts. Furthermore, the current model evaluates objects in isolation, without accounting for intervening obstacles or dynamic movement. Future work aims to extend these static relations into the temporal domain to describe motion and events.
Summary updated