CGGT: Curve-Grounded Geometry Transformer for 3D Parametric Curve Reconstruction
TLDR
CGGT reconstructs editable 3D parametric curves from sparse unposed multi-view images using a curve-grounded transformer and Wireframe-100K dataset.
Reasoning
The paper presents a strong learning-based approach with a novel transformer architecture and a large-scale CAD dataset, showing clear improvements in sparse-view curve reconstruction. However, evaluation appears limited to synthetic renderings, and the scope covers only parametric curves rather than full B-Rep or CSG CAD models.
Read-first score
Read-first score 55.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.
Field roles
Rank sensitivity
Stability: volatile; rank range: 26.
Keyword Scores
Deep Analysis
Innovations
- Introduces CGGT, a Curve-Grounded Geometry Transformer that directly grounds 3D-consistent 2D curve instances from sparse, unposed multi-view images.
- Combines a geometry-aware transformer encoder for multi-view feature learning with a curve-aware masked-attention decoder for cross-view instance association.
- Predicts camera parameters, dense depth maps, and instance-level 2D curve masks in a single forward pass, then lifts them into 3D and refines via fast parametric optimization.
- Introduces Wireframe-100K, a large-scale dataset of 100,000 CAD models with diverse topologies, realistic multi-view renderings, and accurate parametric curve annotations.
- Achieves improved reconstruction accuracy and efficiency, especially in sparse-view settings, and separates persistent 3D structural edges from view-dependent image edges.
Methodology
CGGT uses a geometry-aware transformer encoder for multi-view feature learning and a curve-aware masked-attention decoder for cross-view instance association. In one forward pass it predicts camera parameters, dense depth maps, and instance-level 2D curve masks, which are lifted into 3D and refined through a fast parametric optimization stage. The method is trained on the introduced Wireframe-100K synthetic CAD dataset and evaluated for reconstruction accuracy, efficiency, sparse-view robustness, and real-world generalization.
Key Results
The paper reports substantial improvements in reconstruction accuracy and efficiency over existing NeRF- and 3DGS-based methods, with particular gains under sparse-view conditions and in distinguishing persistent 3D structural edges from view-dependent edges. CGGT also generalizes to real-world images despite being trained only on synthetic data.
Limitations
- Training relies solely on synthetic data, so real-world generalization is reported but not fully characterized in the abstract.