Abstract
Geometric design often relies on hierarchical structures where simple elements are repeated and transformed to create complex wholes. This paper presents the Stochastic Wreath Process, a generative framework that formalizes this hierarchy using the mathematical framework of wreath products. By incorporating noise into these nested transformations, the model can robustly interpret imperfect human sketches. We demonstrate an inverse graphics approach using Approximate Bayesian Computation (ABC) and Reversible Jump Markov Chain Monte Carlo (RJ-MCMC) to recover the underlying structural "DNA" of geometric patterns, ranging from synthetic shapes to real-world architectural floor plans.
Introduction
Human perception naturally organizes geometric patterns into hierarchies of symmetry. A snowflake is seen not as a collection of independent ice crystals, but as a single branch rotated six times. While traditional geometric theories provide a rigorous mathematical language for these structures, they often fail when applied to real-world data. Human drawings are inherently noisy; lines are rarely perfectly straight, and angles seldom meet at exactly 90 degrees.
The challenge lies in bridging the gap between "platonic" mathematical perfection and the "wiggly" reality of human sketches. This research addresses this by treating vision as the inverse of drawing—a process of program synthesis where the goal is to find the simplest hierarchical transformation "program" that explains an observed image.
Methodology: The Stochastic Wreath Process
The core innovation is the Stochastic Wreath Process. It utilizes the wreath product, a group-theoretic method for nesting transformations, to define a generative prior for shapes. In this model, a shape is constructed by taking a base "fiber" (a primitive building block), replicating it, and applying a "control" transformation such as rotation or translation.
To account for human variability, the authors inject a noise parameter $\epsilon$ at every level of the hierarchy. This ensures that a small deviation in a high-level transformation (e.g., a large rotation) correctly propagates its impact down to all constituent parts.
Inverse Graphics via ABC and RJ-MCMC
Recovering the generative history of a shape is framed as an inference problem. Because calculating the exact likelihood of pixel arrangements is computationally intractable, the authors employ Approximate Bayesian Computation (ABC). A stochastic renderer generates candidate shapes based on proposed hierarchies, which are then compared against the target image.
To navigate the vast space of possible nested structures, the model uses Reversible Jump Markov Chain Monte Carlo (RJ-MCMC). This allows the inference engine to "jump" between different levels of complexity—for instance, testing whether a shape is better explained by 4-fold or 5-fold symmetry.
Experimental Evaluation
The model was evaluated across three primary datasets: synthetic ground truth shapes, a prior-sampled test set of 50 examples (including hand-drawn versions), and real-world architectural sketches of iconic buildings like the Dome of the Rock and Villa La Rotonda.
Key Results
The model demonstrated significant robustness in structural recovery:
- Recoverability: On exact computer-rendered shapes, the model achieved 42.76% full recoverability of the underlying group structure. When relaxed to "recoverability up to occupancy," the success rate improved to 55.84%.
- Hand-Drawn Robustness: For noisy, hand-drawn sketches, the model maintained a 52.28% success rate, validating the effectiveness of the stochastic noise modeling.
- Generative Completion: The system can perform "automatic fill-ins." By inferring the underlying symmetry from a partial sketch (e.g., two sides of a hexagon), it can suggest the remaining components of the shape.
Discussion and Future Work
The Stochastic Wreath Process views shape as a "memory store" of generative actions. By maximizing the "transfer" of parts—explaining many lines through a single repeated transformation—the model identifies the most meaningful structural representations.
Current limitations include a restriction to 2D monochromatic sketches and the high computational cost of MCMC inference. Future work aims to extend this framework to 3D and color, and to explore more efficient inference techniques. Beyond geometry, the hierarchical logic of the wreath process holds potential for analyzing other structured domains, such as musical compositions or genetic regulatory networks.
Summary updated