The paper introduces a novel shape representation that models objects as compositions of deformable "lumps of clay"—parameterized superquadrics—combined through Boolean operations and generic formative processes such as stretching, bending, twisting, and tapering. This intermediate‑grain "part‑and‑process" model bridges the gap between low‑level image‑formation models and high‑level specialized CAD models, aiming to be both succinct and perceptually natural. The author argues that existing vision models (point‑wise optics‑based or object‑specific CAD) suffer from fragility, under‑determination, and lack of scalability for general‑purpose robots. By exploiting the dual relationship between superquadric surface normals and shape, the paper derives analytic equations that allow over‑constrained recovery of 3‑D part parameters from noisy image data, providing a built‑in reliability check. Psychophysical experiments and protocol analyses show that humans pre‑attentively detect the same primitive features (convexity, squareness‑roundness, bending, tapering, stretching) that the representation uses. A real‑time interactive modeling system, SuperSketch, demonstrates the practicality and naturalness of the primitives. For highly detailed natural surfaces (mountains, clouds), the author extends the framework with fractal constructions that capture roughness statistically while preserving coarse shape. The paper concludes with successful demonstrations of part segmentation, recovery of deformed primitives, and a discussion of limitations and future extensions (branching structures, particle systems).
Topics
Critique of Existing Vision Models: Image‑Formation vs. Specialized CAD
Analyzes the limitations of low‑level optics‑based models (under‑determination, fragile assumptions) and high‑level CAD models (combinatorial explosion, lack of learning) for general‑purpose robotics.
Part‑and‑Process Paradigm: An Intermediate‑Grain Representation
Proposes a middle‑ground model that uses generic formative processes (bending, twisting, interpenetration) to build objects from superquadric "lumps of clay".
Superquadrics as Primitive Lumps: Mathematics and Deformations
Introduces the superquadric family, its parameterization, and the four deformation operators (stretch, bend, twist, taper) plus Boolean combination to generate complex shapes.
Psychophysical and Protocol Evidence for Human Alignment
Presents experimental findings (Treisman pre‑attentive detection, protocol analysis of verbal descriptions) showing that humans naturally use the same primitive features the model defines.
SuperSketch: Interactive Modeling System Demonstrating Naturalness
Describes the real‑time 3‑D modeling interface built on the representation, highlighting rapid construction of complex scenes and user‑effort reduction.
Fractal Extension for Complex Natural Surfaces
Extends the primitive framework with recursive superquadric sums to model rough natural forms (mountains, clouds) using a roughness ratio r and fractal dimension D.
Duality of Surface Normals and Shape: Overconstrained Recovery
Derives analytic relations between surface normals and superquadric parameters, enabling linear regression and closed‑form solutions for 3‑D shape from tilt τ alone.
Segmentation and Recovery of Part Descriptions from Depth Data
Shows how ratios τ/dτ/dx and τ/dτ/dy produce linearly varying domains that segment images, recover primitive parameters, and handle deformed parts (bent, tapered).