Img2CADSeq: Image-to-CAD Generation via Sequence-Based Diffusion
TLDR
Img2CADSeq uses a multi-stage pipeline with hierarchical codebook and VQ-Diffusion to generate CAD BRep sequences from single-view images, outperforming SOTA on new datasets.
Reasoning
The paper introduces a novel approach with strong empirical results and new datasets, but the abstract lacks discussion of limitations or failure cases. The multi-stage pipeline may introduce complexity, and the reliance on point cloud intermediates could be a bottleneck.
Read-first score
Read-first score 63.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 111.
Field roles
Rank sensitivity
Stability: volatile; rank range: 31.
Keyword Scores
Deep Analysis
Innovations
- Multi-stage pipeline with three-level hierarchical codebook for CAD sequences
- Importance prioritization valuing profiles over details to compress long sequences
- Coarse-to-fine point cloud intermediate to bridge 2D-3D modality gap
- Contrastive learning aligning 2D visual features with 3D CAD sequences for VQ-Diffusion conditioning
- Introduction of CAD-220K and PrintCAD datasets for industrial domain adaptation
- Generation of standard STEP files directly usable in commercial CAD software
Methodology
Img2CADSeq is a multi-stage pipeline that encodes CAD sequences into a three-level hierarchical codebook with importance prioritization, then uses a coarse-to-fine point cloud intermediate and contrastive learning to align 2D image features with 3D CAD sequences, conditioning a VQ-Diffusion model. It is trained on newly introduced CAD-220K and PrintCAD datasets.
Key Results
Img2CADSeq significantly outperforms state-of-the-art methods in image-to-CAD generation, producing standard STEP files that can be directly used in commercial CAD software.