Abstract
Mesh2Fab is a data-driven framework designed to bridge the gap between abstract 3D modeling and material-specific fabrication. Unlike traditional rationalization techniques that seek the closest geometric approximation, Mesh2Fab "reforms" shapes to align with the structural and aesthetic "language" of a target material, such as wood or metal. By leveraging a database of real-world designs, the system intelligently replaces components and optimizes joint configurations to ensure physical manufacturability and visual plausibility.
Introduction
In digital design, geometry is often decoupled from the physical properties of materials. A spindly, curved chair modeled in CAD might look elegant with a wood texture, but its thin members and complex joints would likely fail if constructed from actual timber. Mesh2Fab addresses this by treating material-specific fabrication as a shape-reforming problem. It recognizes that materials like wood favor orthogonal, thick components, whereas metal allows for slender, curved forms.
Methodology
The Mesh2Fab pipeline consists of four primary stages:
- Preprocessing & Analysis: The input mesh is decomposed into functional parts and analyzed for geometric properties like thickness and connectivity.
- Part Replacement: Using a database of 152 annotated models, the system formulates a multi-label assignment problem. It replaces input parts with material-appropriate components using loopy belief propagation to balance shape similarity and contact consistency.
- Structure Optimization: The system nudges parts into more "natural" configurations (e.g., forcing wooden joints toward 90 degrees) based on contact angle histograms derived from the database.
- Joint Inference: Finally, the system predicts specific fabrication joints, such as mortise-and-tenon for wood or welding for metal, based on local geometry.
Results and Evaluation
The framework was evaluated on various man-made models, including chairs, tables, and beds. A user study involving 30 participants demonstrated that the reformed shapes successfully captured material-specific aesthetics, with users correctly identifying the intended material 91.5% of the time based solely on geometry. Furthermore, the system achieved 84.2% accuracy in joint type prediction. The efficiency of the process is notable, with full reformation typically taking around 15 seconds.
Discussion and Limitations
While Mesh2Fab provides a robust method for adapting digital designs to physical reality, it remains dependent on the diversity of its underlying database. If the database lacks geometrically similar parts, results may appear distorted. Additionally, the current system focuses on wood and metal, excluding materials like plastics. Future work could incorporate structural load-bearing analysis and the generation of internal hardware geometry for connectors like screws or rivets.
Summary updated