CGGT:
Curve-Grounded Geometry Transformer for 3D Parametric Curve Reconstruction
Editable 3D curves from sparse, unposed images. CGGT grounds consistent curve instances across views, predicts geometry, and reconstructs compact CAD-style primitives.
01 / OVERVIEW
From pixels to parametric structure
Existing curve reconstruction pipelines often need dense calibrated views, precomputed edge maps, or costly per-scene fitting. CGGT predicts camera parameters, depth, and cross-view curve instances in a single forward pass. A fast fitting stage then lifts these observations into editable 3D primitives.
PAPER VIDEO
Watch the paper overview
A short introduction to CGGT and its curve reconstruction results.
02 / METHOD
Two stages, one coherent wireframe
Geometry-aware multi-view features and curve-aware decoding separate persistent structural curves from appearance edges. Camera and depth predictions support 3D lifting and parametric refinement.
03 / DATASET
Wireframe-100K
A synthetic dataset built from 100,000 CAD models with diverse topologies, multi-view renderings, and parametric curve annotations. The released data-preparation scripts document the ABC rendering and curve metadata pipeline; the dataset files themselves are not hosted in the code repository.
Data preparation guide ↗
04 / RESULTS
Structural curves across settings
Qualitative results from the paper show coherent curve reconstruction on synthetic benchmarks and generalization to real images.
05 / CITATION
Cite this work
@inproceedings{gao2026cggt,
title = {CGGT: Curve-Grounded Geometry Transformer for 3D Parametric Curve Reconstruction},
author = {Gao, Zhirui and Yi, Renjiao and Ye, Yunfan and Hu, Ruizhen and Zhu, Chenyang and Chen, Wei and Xu, Kai},
booktitle = {SIGGRAPH Asia 2026 Conference Papers},
year = {2026},
doi = {10.1145/3829340.3842167}
}