acceptodds
Withdrawn from ICLR 2027

ED2CAD: Parametric CAD Reconstruction from Dimension-Annotated Engineering Drawings

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstructing editable 3D parametric models from dimension-annotated engineering drawings remains a challenging problem. We propose ED2CAD, a dimension-aware framework for reconstructing parametric Computer-Aided Design (CAD) models from multimodal engineering drawings. To support this task, we develop an automated data pipeline that generates paired raster and vector drawings from B-rep geometry, including orthographic views with visible lines, hidden lines, and dimensional annotations. A raster-vector Transformer jointly encodes and fuses multimodal drawing features to autoregressively predict CAD commands and parameters. To improve geometric fidelity, we further introduce a point cloud geometry proxy that provides 3D supervision based on Chamfer distance. The predicted command sequences are instantiated through a CAD kernel to obtain executable 3D models for geometric evaluation. Extensive experiments demonstrate the effectiveness of dimensional annotations, multimodal inputs, geometric supervision, and the proposed architecture.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer-Aided Design, Dimension-Annotated Engineering Drawings, Parametric CAD Reconstruction, Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UgwYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a practically meaningful benchmark for CAD reconstruction and presents a multimodal model for parametric CAD reconstruction from dimension-annotated engineering drawings. The benchmark is based on WHUCAD and aligns tokenized CAD sequences with annotated PNG, SVG and DXF drawings, providing data resources for CAD reconstruction from engineering drawings. The proposed model encodes and fuses multimodal drawing features to autoregressively predict CAD commands and parameters, while a point-cloud geometry proxy provides auxiliary 3D supervision. The predicted sequences are executed through a CAD kernel to obtain 3D models. Also, extensive experiments explained the effectiveness of the model and the proposed modules. The paper evaluates the framework through baseline comparisons and ablation studies on input modalities, dimension annotations, and model components.

Critical strengths
  1. This paper introduces a practically meaningful benchmark for CAD reconstruction. The benchmark is based on WHUCAD and aligns tokenized CAD sequences with annotated PNG, SVG and DXF drawings, providing data resources for CAD reconstruction from engineering drawings.
  2. The paper presents a comprehensive model with extensive experiments. The model incorporates multimodal inputs, a fusion module to combine their features and a point-cloud geometry proxy head to assist reconstruction.
Critical weaknesses
  1. The distinct contribution of the ED2CAD dataset over existing datasets is unclear. ED2CAD derives its data from WHUCAD and its contributions mainly focus on engineering drawings, annotations and their alignment. However, its differences from existing datasets in terms of operation coverage and geometric complexity are insufficiently demonstrated and it mainly appears to provide additional data representations. Therefore, I find it difficult to assess the dataset’s distinct contribution.
  2. The proposed model consists mainly of commonly used building blocks in computer vision, making its methodological novelty questionable. Furthermore, the experiments do not clearly demonstrate the improvement of the full ED2CAD model over a basic version without the proposed modules.
Questions for authors
  1. What distinguishes ED2CAD from existing datasets?
    Could the authors provide a systematic comparison with closely related datasets, regarding operation coverage, geometric complexity, and dimension-annotation coverage? Which research capabilities does ED2CAD enable that are not adequately supported by these existing resources?
  2. How is the quality of the generated data verified?
    Could the authors report how frequently basename-based matching, projection fallback, skipped edges, and annotation collision overrides occur? Have manual inspections or other checks been conducted to assess pairing correctness, drawing completeness, and dimension-annotation accuracy?
  3. What is the overall gain from the proposed components over a basic architecture? Could the authors report a basic configuration that uses direct fusion without fusion-stage modality embeddings or the geometry proxy, under the same inputs and training settings? Comparing this configuration with the full model would help establish the overall benefit of the proposed additions.
Reviewer_u1BZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ED2CAD, a framework for reconstructing parametric CAD command sequences from dimension-annotated engineering drawings. The authors build a dataset of 133,656 samples by generating aligned PNG, DXF, and SVG drawings from WHUCAD models. The model uses separate raster and vector encoders, a fusion Transformer, and an autoregressive CAD decoder. An auxiliary point-cloud prediction head provides additional geometric supervision.

Critical strengths

The data construction pipeline is a useful contribution. Reconstructing editable CAD models from engineering drawings is a useful problem. The dataset construction involves more than rendering existing models: it includes visibility analysis, dimension selection and placement, and alignment between drawings and CAD sequences.

Critical weaknesses

The methodological contribution appears limited. The architecture mainly combines modality-specific Transformer encoders, token fusion, autoregressive sequence prediction, and an auxiliary point-cloud loss. I see substantial work in the data pipeline, but the model itself reads as an incremental extension of existing CAD sequence predictors.

The ablations do not separate multimodal benefits from additional model capacity. The paper attributes the improvements in Table 1 to complementary information across PNG, DXF, and SVG. However, enabling more modalities also activates additional independently parameterized encoders and increases the number of tokens processed by the fusion module and decoder.

Questions for authors

How does the decoder preserve the identities of previously generated arguments?

How dependent are the results on command-assisted annotation generation?

Reviewer_EQ9yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ED2CAD for reconstructing parametric CAD modeling sequences from dimension-annotated engineering drawings. It introduces an automated data annotation pipeline (covering raster and vector drawings, visible and hidden lines, and dimension annotations), a raster-vector fusion transformer, and a geometry proxy for auxiliary 3D supervision. Experiments on WHUCAD and DeepCAD demonstrate improvements over existing baselines.

Critical strengths
  1. It extends existing engineering-drawing annotation pipelines by providing a practical method for automatically generating dimension annotations.
  2. The geometry proxy provides auxiliary 3D geometric supervision during training, with ablation results supporting its effectiveness in improving CAD reconstruction performance.
Critical weaknesses
  1. Although the ablation studies in Table 1 show improvements from raster-vector fusion, it remains unclear whether these gains reflect genuinely complementary information across modalities or simply better fitting to the current data distributions through additional inputs and increased model capacity. Cross-distribution evaluation would provide stronger evidence of the generalizability of these gains.
  2. The paper specifies fixed orthographic projection directions, but it does not clarify whether objects are aligned to a standard orientation. Different orientations of the same object can result in different three-view drawings, yet the paper does not include any experiments to evaluate whether the method remains effective under such variations.
Questions for authors
  1. Could the authors provide cross-distribution evaluation results to show whether the fusion gains generalize beyond the current data distributions? Additionally, given that SVG and DXF essentially represent the same vector geometry, why is it necessary to use both?
  2. Could the authors clarify whether objects are aligned to a canonical orientation before generating the fixed orthographic views? If not, is there evidence the method remains effective under different orientations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.