CAD-SIGNet: CAD Language Inference from Point Clouds Using Layer-Wise Sketch Instance Guided Attention
TLDR
End-to-end auto-regressive model that recovers CAD design history (sketch-and-extrusion) from point clouds using sketch instance guided attention.
Reasoning
The paper introduces a novel attention module for fine-grained sketch reconstruction and supports multiple plausible design histories, which are strengths. However, it is limited to sketch-and-extrusion operations and does not address other CAD primitives or real-world scan noise, and the evaluation is only on public datasets.
Read-first score
Read-first score 56.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 88.
Field roles
Rank sensitivity
Stability: volatile; rank range: 35.
Keyword Scores
Deep Analysis
Innovations
- End-to-end trainable autoregressive architecture for CAD design history recovery from point clouds
- Layer-wise cross-attention between point cloud and CAD language embedding
- Sketch Instance Guided Attention (SGA) module for fine-grained sketch detail reconstruction
- Generation of multiple plausible design choices enabling interactive reverse engineering
Methodology
CAD-SIGNet is an autoregressive model that recovers a sequence of sketch-and-extrusion operations from an input point cloud. It employs layer-wise cross-attention between point cloud features and CAD language embeddings, with a Sketch Instance Guided Attention module to reconstruct fine sketch details. The model is trained end-to-end on public CAD datasets and evaluated on full design history recovery and conditional auto-completion tasks.
Key Results
CAD-SIGNet outperforms existing baselines on full design history recovery and conditional auto-completion from point clouds, as shown by extensive experiments on publicly available CAD datasets.