Year: 2025
Type: article
Abstract: Reconstructing computer-aided design (CAD) models from geometric models has long been a fundamental yet challenging research problem. In this work, we propose a novel contrastive learning framework, CADCL, which reconstructs parametric CAD sequences from B-rep models. The framework consists of two stages. In the first stage, a transformer-based module is trained to encode CAD sequences into latent embeddings. In the second stage, the input B-rep is represented as a graph and jointly process with the CAD sequence embeddings obtained from the first stage. The final output is a parametric CAD sequence. Different from existing approaches, this paper innovatively incubates a contrastive learning approach for B-rep embeddings and CAD sequence embeddings, which enables the B-rep embeddings to effectively capture information align with the parametric CAD structure. In this way, the B-rep features can be more accurately decoded into CAD sequences. Extensive experimental results on both the simple DeepCAD dataset and advanced WHUCAD dataset demonstrate that our method outperforms existing approaches, and the generated CAD sequences can be successfully imported and edited in standard CAD modeling software.
DOI: 10.1093/jcde/qwaf102
Authors: JianFei Liang, FaZhi He, RuBin Fan, YiYang Chu, Xiaohu Yan
Institutions: Wuhan University, Shenzhen Polytechnic University
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,650
Awards: G2733791690, 复杂合作势态下协同设计数据保障方法研究
OpenAlex ID: W4414793032