SECAD-Net: Self-Supervised CAD Reconstruction by Learning Sketch-Extrude Operations
TLDR
SECAD-Net self-supervisedly reconstructs editable CAD models by learning sketch-extrude operations from raw geometry using implicit fields.
Reasoning
The paper introduces a novel self-supervised approach for CAD reconstruction that leverages sketch-extrude operations, showing strong results on standard datasets and enabling editing. However, it is limited to extrusion-based modeling and does not address other CAD operations or complex topologies.
Read-first score
Read-first score 55.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 88.
Field roles
Rank sensitivity
Stability: volatile; rank range: 49.
Keyword Scores
Deep Analysis
Innovations
- Self-supervised learning for CAD reconstruction without requiring labels
- Reconstruction via learning 2D sketches and 3D extrusion parameters, mimicking modern CAD modeling language
- Implicit field representation for sketches enabling latent space interpolation and CAD variations
- Combining extrusion cylinders with Boolean union to approximate target geometry
Methodology
SECADNet is an end-to-end neural network that learns to reconstruct CAD models from raw geometry by predicting 2D sketches and 3D extrusion parameters in a self-supervised manner. It generates extrusion cylinders by extruding each sketch from a 2D plane into a 3D body and combines them via Boolean union to approximate the target shape. Implicit fields are used for sketch representation, allowing interpolation in the sketch latent space for CAD variations.
Key Results
Experiments on ABC and Fusion 360 datasets demonstrate that SECADNet outperforms state-of-the-art alternatives, including a closely related supervised method, and is applicable to CAD editing and single-view reconstruction.