Accepted at ECCV'24
3D reverse engineering, in which a CAD model is inferred given a 3D scan of a physical object, is a research direction that offers many promising practical applications. This paper proposes TransCAD, an end-to-end transformer-based architecture that predicts the CAD sequence from a point cloud. TransCAD leverages the structure of CAD sequences by using a hierarchical learning strategy. A loop refiner is also introduced to regress sketch primitive parameters. Rigorous experimentation on the DeepCAD and Fusion360 datasets show that TransCAD achieves state-of-the-art results. The result analysis is supported with a proposed metric for CAD sequence, the mean Average Precision of CAD Sequence, that addresses the limitations of existing metrics.
1) We propose TransCAD, a novel hierarchical architecture for feature-based reverse engineering. Our model is single-stage and end-to-end trainable. TransCAD allows for a compact CAD sequence representation that does not include categorical types and enables cascaded coordinate refinement.
2) We identify several limitations of the existing evaluation framework for feature-based reverse engineering and propose a new evaluation metric framework to ensure fair comparison among diverse network architectures.
3) Our model surpasses the performance of recent generative-based approaches while also bridging the gap to real-world applications by exhibiting robustness to perturbed point clouds.
The present project is supported by the National Research Fund, Luxembourg under the BRIDGES2021/IS/16849599/FREE-3D and IF/17052459/CASCADES projects, and by Artec 3D.
@misc{dupont2024transcad,
title={TransCAD: A Hierarchical Transformer for CAD Sequence Inference from Point Clouds},
author={Elona Dupont and Kseniya Cherenkova and Dimitrios Mallis and Gleb Gusev and Anis Kacem and Djamila Aouada},
year={2024},
eprint={2407.12702},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2407.12702},
}