Introduction
Recognizing three-dimensional objects from two-dimensional images remains a fundamental challenge in computer vision. In their 1980 work, Shapiro et al. proposed a shift from high-fidelity geometric models to "rough" relational models. They argued that because real-world image data is often noisy, distorted, or poorly segmented, the models used for classification should prioritize "gross-level" topology over fine geometric detail. This approach bridges the gap between segmented 2D images and high-level 3D categories, such as identifying a chair without requiring perfect geometric reconstruction.
The Three Primitives
The framework represents complex 3D objects using three basic primitives based on their significant dimensions:
- Sticks: Parts where only one dimension is significant, typically long and thin.
- Plates: Flat parts with two significant dimensions, representing surfaces bounded by thin edges.
- Blobs: Parts where all three dimensions are significant, being neither thin nor flat.
These primitives are defined as "near-convex," ensuring that sticks do not bend excessively and blobs lack deep concavities. This constraint simplifies the computational representation while maintaining enough descriptive power to model most common man-made objects.
Relational Modeling and Matching
The core of this approach is the relational description. An object is defined not just by its constituent parts, but by the CONNECTS/SUPPORTS relations between them. This includes specifying connection types (e.g., end-to-end, center-to-center) and permissible angle ranges rather than exact geometric values. This flexibility allows the system to account for the variability inherent in 2D projections and perspective distortion.
The recognition process involves three main steps:
- Decomposition: A 2D view is clustered into simple parts using graph-theoretic methods.
- Screening: An Attribute/Value (A/V) table, storing global features like the number of supports or the type of the topmost part, narrows the search space in the model database.
- Inexact Matching: The system performs relational matching to map 2D segments to 3D model parts, minimizing the "error" resulting from unmet constraints.
Conclusion
By focusing on topology over strict geometry, the "Sticks, Plates, and Blobs" method provides a robust framework for object recognition in imperfect conditions. While challenges like perspective distortion and support inference remain, the use of hierarchical filtering and inexact matching established a foundational strategy for high-level scene analysis that is resilient to sensor noise and segmentation errors.
Summary updated