CAD-VGDrawing Dataset Overview

• CAD-VGDrawing is a benchmark dataset that pairs 2D vector drawings with parametric CAD models to enable automated reconstruction from detailed annotations.
• It employs a dual-decoder transformer architecture to predict both symbolic construction commands and numerical geometric parameters accurately.
• The dataset facilitates research in generative CAD modeling, representation learning, and automated design workflows by providing a structured, sequence-to-sequence learning framework.

The CAD-VGDrawing Dataset is a structured benchmark for bridging engineering vector graphic drawings and parametric CAD model generation. Developed to support sequence-to-sequence learning for CAD construction from vectorized 2D inputs, it provides paired examples connecting vector graphical primitives with parametric CAD commands and geometric parameters. The dataset enables both symbolic command prediction and precise geometric estimation, thus facilitating direct, automated translation from drawings to editable CAD models. It is intended for research into generative modeling, representation learning, and automated design workflows.

1. Dataset Structure and Composition

The CAD-VGDrawing dataset consists of paired samples, each containing:

• A vectorized engineering drawing (such as SVG or similar, representing lines, arcs, curves, etc.)
• The corresponding parametric CAD model, represented as a sequence of construction commands with associated geometric parameters

Drawings are sourced from technically precise engineering sketches and vector graphics, including freehand or technical drawing styles. On the CAD side, parametric models are collected from open-source communities and industry repositories, using tools such as FreeCAD and PythonOCC (Qin et al., 26 Aug 2025). Each paired sample encapsulates the complete transformation from vector graphic primitives to an explicit, editable CAD object.

2. Data Annotation and Pairing Protocol

Annotation is performed by constructing a detailed correspondence between vector primitives and CAD model operations:

• Each drawing primitive (e.g., line, arc) is mapped to a specific step in the CAD model’s construction
• CAD command sequences are defined with explicit operation labels (e.g., “extrude”, “fillet”, “sketch”) and the associated parameters (lengths, radii, angles, coordinates)
• The process yields a canonical “construction sequence annotation,” facilitating model training for both command order and geometric value prediction

This pairing enables an unambiguous training signal for models performing sequence-to-sequence learning. The framework ensures that both symbolic and geometric aspects of the engineering intent are preserved in the annotation.

3. Processing and Representation of Vector Primitives

To create inputs suitable for neural sequence models, each drawing’s vector elements are:

• Normalized and canonicalized into a tokenized sequence, encoding both the type and required geometric data for each primitive
• Represented in a manner invariant to typical drawing variations, ensuring robustness across diverse samples and drawing habits

This “network-friendly” format ensures efficient transformer-based encoding, capturing both global layout and local feature structure essential for downstream CAD reconstruction.

Primitive Type	Token Format	Parameters Included
Line	LINE	start/end coordinates
Arc	ARC	center, radius, angles
Curve	CURVE	control points, type

4. Sequence-to-Sequence Learning Framework

The principal technical approach for CAD-VGDrawing is a dual-decoder transformer architecture:

• Encoder: Processes the vector primitive sequence, extracting geometric and structural context
• Command Decoder: Predicts the CAD construction command sequence (“extrude”, “sketch”, etc.)
• Parameter Decoder: Simultaneously estimates the numeric geometric parameters relevant to each command

Formally, given input sequence $X$ (vector primitives) and output sequence $Y = \{(c_1, p_1), (c_2, p_2), ..., (c_T, p_T)\}$, the conditional probability is factorized as:

$$P(Y|X) = \prod_t P(c_t|X,c_{<t}) \cdot P(p_t|X,c_{\leq t})$$

This structure decouples symbolic operation prediction from numeric geometric regression, yielding models capable of reconstructing valid, editable CAD structures from ambiguous or noisy vector drawings.

The framework introduces a soft target distribution loss:

$L_{soft} = -\sum_k q_k \log p_k$

with $q_k$ as the smoothed ground-truth token probability and $p_k$ as the predicted probability. This loss is applied to both command and parameter branches, mitigating over-confident or unstable predictions and accommodating inherent ambiguity in parametric values.

5. Experimental Results and Evaluation Metrics

Experiments on the CAD-VGDrawing dataset employ metrics tailored to both symbolic and geometric aspects of CAD reconstruction:

• Reconstruction Accuracy: Measures the proportion of CAD command sequences that match ground-truth construction order
• Sequence Validity: Fraction of generated sequences resulting in valid, editable CAD models
• Parameter Error: Quantifies deviations of predicted geometric parameters from annotated ground-truth values

Empirically, the dual-decoder transformer paired with the soft target loss outperforms baseline approaches, achieving higher fidelity in both command prediction and parameter estimation. Models demonstrate effective end-to-end transformation capability from raw vector drawings to usable CAD objects, supporting further engineering automation efforts (Qin et al., 26 Aug 2025).

6. Research Applications and Implications

The CAD-VGDrawing dataset serves as an enabling tool for a wide range of academic and industrial research agendas, including:

• Automating digitalization of hand-drawn sketches into parametric CAD models
• Embedding generative modeling in engineering workflows, allowing rapid transition from conceptual sketches to precise designs
• Supporting multi-modal learning where graphical inputs are paired with structured, symbolic engineering outputs
• Facilitating studies in design re-use, assembly retrieval, and interactive engineering search engines

The structure and annotation protocol of CAD-VGDrawing directly support future work in multi-modal generative modeling and automated design workflows. It lays foundational groundwork for integrating advanced neural sequence models into established engineering design processes.

7. Dataset Access, Licensing, and Future Directions

The CAD-VGDrawing dataset is made available to the research community, with code and data accessible at https://github.com/lllssc/Drawing2CAD (Qin et al., 26 Aug 2025). Licensing follows standard academic practices, typically under open or research-oriented terms such as Creative Commons or similar. Usage instructions and further details are provided in the supplementary materials and project website.

A plausible implication is that future work may extend the dataset’s coverage to include further drawing styles, additional parametric CAD operations, and more varied annotation protocols. Potential directions include enhancing the editability of generated models, integrating richer vector graphic document features, and expanding to support assembly or multi-part parametric modeling.

The CAD-VGDrawing dataset represents a significant step in unifying the domains of vector graphic representation and precise engineering design, directly supporting new research and applications in CAD generative modeling, retrieval, and workflow automation.