Sage Journals: Discover world-class research

Abstract

3D shape learning is an important research topic in computer vision, in which the datasets play a critical role. However, most of the existing 3D datasets use voxels, point clouds, mesh, and B-rep, which are not parametric and feature-based. Thus they can not support the generation of real-world engineering computer-aided design (CAD) models with complicated shape features. Furthermore, they are based on 3D geometry results without human-computer interaction (HCI) history. This work is the first to provide a full parametric and feature-based CAD dataset with a selection mechanism to support HCI in 3D learning. First, unlike existing datasets, mainly composed of simple features (typical sketch and extrude), we devise complicated engineering features, such as fillet, chamfer, mirror, pocket, groove, and revolve. Second, different from the monotonous combination of features, we invent a select mechanism to mimic how human focuses on and selects a particular topological entity. The proposed mechanism establishes the relationships among complicated engineering features, which fully express the design intention and design knowledge of human CAD engineers. Therefore, it can process advanced 3D features for real-world engineering shapes. The experiments show that the proposed dataset outperforms existing CAD datasets in both reconstruction and generation tasks. In quantitative experiment, the proposed dataset demonstrates better prediction accuracy than other parametric datasets. Furthermore, CAD models generated from the proposed dataset comply with semantics of the human CAD engineers and can be edited and redesigned via mainstream industrial CAD software.

Keywords

Computer-aided design human-computer interaction parametric and historical feature-based advanced features selection mechanism

1. Introduction

Computer-aided design (CAD) plays an important role in modern engineering and manufacturing [1, 2]. Parametric CAD models are particularly important in the design process. Design semantics and human-computer interaction (HCI) represent the design intent and design knowledge of human CAD engineers. Adding HCI metadata to CAD files creates more possibilities for customized manufacturing [3, 4]. However, existing 3D geometric data do not possess these characteristics.

With the drastic development of deep learning for computer science [5, 6, 7, 8, 9, 10, 11, 12], how to learn 3D shapes is becoming an important research area in which 3D datasets play a critical role [13, 14, 15, 16]. Typical 3D datasets, such as voxel-based, point cloud-based, and mesh-based datasets, are typically used in classification tasks, segmentation tasks, and retrieval tasks [17, 18, 19, 20, 21, 22]. However, these datasets are geometry results of 3D shapes, which do not contain HCI processes or parametric features.

Some recent works have begun to consider CAD processing as a sequence of CAD commands, in which the CAD datasets typically include simple features and monotonic combinations of simple features, such as sketch-extrude to learn 3D modeling [23, 24]. Unfortunately, these works continue to ignore the HCI processing of modern CAD systems, which inherently adopts human-centered computing architecture and parametric feature-based history to express the design intention and design knowledge of human CAD engineers. Thus, the above datasets and learning approaches are still far from real-world engineering applications.

To this end, this paper is the first to provide a full parametric and feature-based CAD dataset (WHUCAD1

¹
We will open source all of the WHUCAD dataset and codes at https://github.com/fazhihe/WHUCAD.
) to support HCI in 3D learning. This dataset can generate CAD models with real-world engineering features as advanced as those created by human engineers. The WHUCAD provides 146,066 CAD models covering 18 types of features.

The main contributions of this paper are as follows:

•
First, unlike the previous DeepCAD dataset, in which the data size is large but the feature types are simple, WHUCAD includes industrial-level features such as extrude, fillet, chamfer, revolve, groove, pocket, shell, mirror, draft, and hole. There are 4 times more feature types in the WHUCAD dataset than in the DeepCAD dataset and the Fusion 360 Gallery dataset, as shown in Table 1.
•
Second, beyond the existing dataset, in which the feature history is a monotonic combination of simple features, WHUCAD devises a learning-based HCI approach, i.e., a selection mechanism, which enables a deep learning network to focus on a particular topological entity. The proposed HCI mechanism learns how a human engineer establishes the relationship between complicated engineering features and topological entities and therefore can fully express the design intentions and design knowledge of human-like AI CAD engineers in the process of 3D learning.
•
Third, CAD models generated from the proposed network trained by WHUCAD can be easily edited and redesigned on mainstream industrial CAD software by human engineers. Therefore, our dataset and discovery are nontrivial and can be a cornerstone for various real-world engineering tasks involving parametric and historical feature-based CAD.

This paper is organized as follows. Section 2 introduces related works on 3D datasets. Section 3 describes WHUCAD in detail. Section 4 shows the results of the experiments and comparisons. Section 7 concludes the paper, including suggestions for future work.
2. Related work

Currently, artificial intelligence (AI) and machine learning impact computer science and HCIs to a large degree [25, 26]. This paper presents a full parametric and feature-based CAD dataset (WHUCAD) to support HCI in 3D learning. Related works are cataloged in the following subsections.

2.1 B-rep-based CAD dataset

B-rep is the mathematical representation of the 3D shape. B-rep-based CAD models are geometric models (mathematically defined geometry and topology) that can be further discretized into point clouds and meshes.

The ShapeNet dataset is a richly annotated large-scale repository of shapes [27, 28], which mainly includes point clouds and meshes for deep learning [29, 30, 31, 32, 33, 34, 35, 36].

Many other 3D datasets are the same as ShapeNet [37, 38, 39]. For example, ModelNet [40] is a dataset containing 151,128 models belonging to 660 unique object categories. It consists of the point cloud, mesh, and B-rep [41, 42].

For another example, the ABC dataset is a collection of one million 3D models proposed by Koch in 2019 [43]. The models are collected from Onshape,2

²
https://www.onshape.com/en/.
an online CAD platform. ABC primarily provides B-rep data as well as point clouds, meshes, and images [44, 45, 46, 47, 48, 36, 49]. Since the B-rep is the mathematical representation of a 3D shape, the normal of the 3D shape can be calculated as the ground truth [50]. Therefore, in the original ABC paper, the authors use the ground truth to verify 3D deep learning networks [51, 52, 53]. Additionally, due to the mathematical presentation of B-rep, the ABC dataset can resample the surface with arbitrary resolutions for different tasks.

The above datasets are geometric results of CAD models but not parametric CAD models.
2.2 Modern CAD systems with human-centered computing architecture and parametric ability

Modern CAD systems adopt human-centered computing architecture and parametric feature-based history to express the design intention and design knowledge of human CAD engineers [54].

Figure 1.

In the process of human-computer interaction, a CAD user is building a CAD model. First, the user needs to conceptualize the model and outline the steps required for creation. After building a generally complete model, the user needs to select specific topology elements to fine-tune or operate advanced features. This operation may be repeated several times. These HCI processes reflect the design intent and design knowledge of the human CAD engineer.

Overall, modern CAD systems include two types of model representations. The first one is the B-rep, which is the geometric model and can be further discretized into a point cloud and mesh. The second is the parametric and historical feature-based model (modeling history), which is a set of modeling operations (features) that are issued by a human engineer during the modeling process. The parametric models fully represent the design intention and design knowledge of a human engineer.

The B-rep is the final result (similar to a snapshot) of the parametric model. Thus, one parametric CAD model can derive a family of different B-rep models with different parametric values and sizes [55, 56].

Therefore, for parametric and historical feature-based CAD systems, the parametric model is the source model, which is created by a human engineer with many analyses and considerable labor, while the B-rep is the resulting model (derived from the parametric model) [57]. When a parametric model is converted into a B-rep model, the design intention, design history, and design knowledge are lost [58, 59]. Similar to the software engineering field, where “source code” is more valuable than “executable code”, in the CAD field, the “parametric model” is more valuable than the “B-rep model”.

More importantly, as shown in Fig. 1, the parametric and feature-based history contains the HCI process (the selection mechanism). In this way, human engineers can express design intentions and design knowledge with advanced and complicated features.

2.3 CAD dataset considering CAD as a sequence of CAD commands

Some emerging works such as the DeepCAD dataset [23] and the Fusion 360 Gallery dataset [24] consider CAD processing as a sequence of CAD commands.

The DeepCAD dataset is derived from Onshape. The difference is that the DeepCAD dataset improves the ABC dataset by using a sequence of CAD commands to describe 3D shapes. However, the DeepCAD dataset only consists of very simple sketch-extrude features. Due to its lack of the HCI mechanism (the selection feature), the DeepCAD dataset cannot support tasks with advanced features [60, 61, 62]. For example, SkexGen uses the DeepCAD dataset to generate parametric CAD models but contains only sketch-and-extrude modeling features [63]. In addition, the feature history of the DeepCAD dataset is a monotonic combination of simple features, which cannot establish the relationship between complicated engineering features and topological entities. Finally, the majority of models in the DeepCAD dataset are not semantically realistic CAD models, which does not comply with the human thinking of CAD engineers.

There are three versions of the Fusion 360 Gallery dataset: segmentation (B-Rep, mesh, and point cloud) [64], reconstruction (a sequence of CAD commands, such as DeepCAD) [24], and assembly (B-rep and mesh) [65]. Here, we discuss the Fusion 360 Gallery reconstruction version because it considers CAD processing as a sequence of CAD commands. Like DeepCAD, Fusion 360 Gallery consists of only CAD modeling commands with simple modeling features without the HCI mechanism, such as sketch and extrude [24, 66, 61, 62]. Point2Cyl uses the Fusion 360 Gallery to generate a set of extrusion cylinders [67] that do not contain complicated features. The difference is that Fusion 360 Gallery has 8,625 models, while DeepCAD has 178,238 models.

In conclusion, DeepCAD and Fusion 360 Gallery bypass human-computer interactions and therefore do not support advanced features.

To the best of our knowledge, WHUCAD is the first dataset to support human-computer interactions in 3D learning at a high level, similar to mainstream industrial CAD systems; therefore, WHUCAD can efficiently reconstruct and generate CAD models with advanced features for real-world engineering tasks.

Table 1
Differences of features among WHUCAD, DeepCAD and Fusion 360 Gallery

Simple features	Parameters	Interpretation of parameters	DeepCAD	Fusion 360	WHUCAD
Line	$x, y$ :	Line endpoint	✓	✓	✓
Arc	$x, y$ :	Arc endpoint	✓	✓	✓
	$\alpha$ :	Sweep angle
	$f$ :	Counter-clockwise flag
Circle	$x, y$ :	Center	✓	✓	✓
	$r$ :	Radius
Spline		$\Phi$			✓
SCP	$x, y$ :	Spline control point			✓
Extrude	$\theta,\phi,\gamma$ :	Sketch plane orientation	✓	✓	✓
	$p_{x},p_{y},p_{z}$ :	Sketch plane origin
	$s$ :	Scale of associated sketch profile
	$e_{1},e_{2}$ :	Extrude distances toward both sides
	$b$ :	Boolean type
	$u_{1},u_{2}$ :	Extrude type of two sides
Advanced feature	Parameters	Interpretation of parameters	DeepCAD	Fusion 360	WHUCAD
Selection	type:	Type of the selected object			✓
	entity_type:	Type of entity which the selected object is in
	entity_no:	Serial number of the entity above
	no:	Serial number of the selected object
Advanced features	Parameters	Interpretation of parameters	DeepCAD	Fusion 360	WHUCAD
depend on selection
Revolve	$\theta,\phi,\gamma$ :	Sketch plane orientation			✓
	$p_{x},p_{y},p_{z}$ :	Sketch plane origin
	$s$ :	Scale of associated sketch profile
	$e_{1},e_{2}$ :	Revolve angle toward both sides
	$b$ :	Boolean type
Pocket	$\theta,\phi,\gamma$ :	Sketch plane orientation			✓
	$p_{x},p_{y},p_{z}$ :	Sketch plane origin
	$s$ :	Scale of associated sketch profile
	$e_{1},e_{2}$ :	Extrude distances toward both sides
	$u_{1},u_{2}$ :	Extrude type of two sides
Groove	$\theta,\phi,\gamma$ :	Sketch plane orientation			✓
	$p_{x},p_{y},p_{z}$ :	Sketch plane origin
	$s$ :	Scale of associated sketch profile
	$e_{1},e_{2}$ :	Revolve angle toward both sides
Topo		$\Phi$			✓
Shell	$t_{1},t_{2}$ :	Internal thickness and external thickness			✓
Chamfer	$e_{1},e_{2}$ :	Two length of chamfer			✓
Fillet	$r$ :	Radius of fillet			✓
Hole	$x, y, z$ :	Absolute position of center			✓
	$r$ :	Radius
	depth:	Hole depth
	type:	Type of hole
Mirror		$\Phi$			✓
Mirror topo		$\Phi$			✓
Draft	angle:	Angle of draft			✓

3. WHUCAD dataset

Unlike traditional 3D datasets consisting of geometric data, the WHUCAD dataset includes complicated advanced features. More importantly, to support HCI in 3D learning, we devise a selection mechanism for the WHUCAD dataset.

Figure 2 shows examples of the final results in WHUCAD, which show better qualitative comparison results than those of DeepCAD (in Fig. 3(a)) and Fusion 360 Gallery (in Fig. 3(b)).

Figure 2.

A gallery of WHUCAD dataset.

Figure 3.

A gallery of (a) DeepCAD and (b) Fusion 360.

As shown in Figs 2 and 3, the CAD models in the WHUCAD dataset are obviously more complicated and have more parametric and feature-based characteristics than those in the DeepCAD dataset and the Fusion 360 Gallery dataset. Both the DeepCAD dataset and the Fusion 360 Gallery dataset consist of simple features such as sketch and extrude. As a result, the DeepCAD dataset and the Fusion 360 Gallery dataset cannot support the reconstruction and generation of CAD models that contain advanced features, while the WHUCAD dataset can reconstruct and generate complicated 3D shapes with advanced feature operations such as fillet, chamfer, revolve, pocket, groove, shell, hole, mirror, draft, and especially the HCI mechanism (selection feature).

Therefore, the most important value of the WHUCAD dataset is not the data size but the HCI mechanism and feature types, the number of which is 4 times greater than that of the DeepCAD dataset and the Fusion 360 Gallery datasets (further discussed in Section 3.3).

3.1 Parametric and historical feature-based CAD modeling with HCI

In modern industrial CAD/CAE/CAM environments, major 3D shape models can be cataloged as discrete representations, B-rep, and parametric representations [68, 69, 70]. Typical discrete representations are mesh, point cloud, and voxel, which can be converted from CAD models for visualization, CAE analysis, and CAM manufacturing. B-rep represents a solid object by its boundary surfaces, such as faces, edges, and vertices. A parametric model is a set of modeling functions (features) that a human engineer issues during a feature-based modeling history. The parametric model is the source model because the B-rep is the one-case final result (similar to a snapshot) of the parametric model. One parametric model could produce a family of B-rep models as desired by adjusting different parametric values and sizes of the parametric model.

The process of constructing a parametric model is shown in Fig. 4. The fundamental definitions of parametric and historical feature-based CAD models are the following [59, 71].

Definition 1: Parametric and historical feature-based model. The parametric model consists of a history of feature operations (such as extrusion, pocket, fillet, chamfer, shell, and selection) as follows:

\displaystyle M=\bigcup{{}_{i=1}^{n}}f_{i}

(1)

where $M$ describes a parametric and historical feature-based CAD model and $f_{i}$ is the $i^{\text{th}}$ feature operation inside the history of $M$ .

Definition 2: Feature operation. A feature operation consists of feature ontology and corresponding feature parameters as follows:

\displaystyle f=\left\{\bigcup{{}_{i=1}^{n}}S_{i},p\right\}

(2)

where $S_{i}$ is the $i^{th}$ feature ontology in $f$ and $p$ denotes the corresponding parameters related to $f$ .

When an engineer watches a 3D shape, he or she can build a parametric CAD model step by step as shown in Fig. 5.

Sketch is usually the first feature operation for parametric and feature-based 3D modeling. A human engineer issues a sketch feature with some basic shape elements such as lines, arcs, circles, and nonuniform rational B-Splines (NURBS) curves [72, 73].

Then, the extrude feature is issued to extend the 2D sketch into a 3D solid geometry.

For advanced features, HCI is necessary. Therefore, a selection operation is needed. For example, the engineer could select which acute edge (a typical topological entity, as shown in the red acute edge in Fig. 5) will be processed in the next feature operation.

After selecting the acute edges that are watched and focused on by the human engineer, he or she issues a fillet feature operation on the selected edges.

Finally, a smooth 3D shape with a beautiful round edge is generated for real-world engineering applications, that is, a semantically meaningful 3D shape according to the human vision of engineers.

Step 3 and 4 can be iterated for many other complicated features, such as selection-chamfer, selection-revolve, selection-shell, and selection-mirror. A typical revolve feature is shown in Fig. 6, and a typical mirror feature is shown in Fig. 7.

3.2 HCI extended feature sequence in WHUCAD

In WHUCAD, 3D CAD data are described as a sequence of features. As a full parametric and feature-based CAD dataset to support HCI, the representation of the feature parameters of a model in the WHUCAD dataset is as follows:

\displaystyle M=(F_{1},F_{2},\ldots,F_{n})

(3)

where:

\displaystyle F_{i}=(t_{i},p_{i})

(4)

and $t_{i}$ indicates the type of command selected from the elements of the “features” column in Table 2.

Figure 4.

The process of building a parametric CAD model.

Figure 5.

An example of parametric and feature-based modeling.

Figure 6.

An example of the revolve feature.

Table 2

The relationship between the digit of feature vector and the definition of parameters

Features	$x$	$y$	$\alpha$	$f$	$r$	$\theta$	$\phi$	$\gamma$	$p_{x}$	$p_{y}$	$p_{z}$	$s$	$e_{1}$	$e_{2}$	$et_{1}$	$et_{2}$	$a_{1}$	$a_{2}$	$b$	$t_{i}$	$t_{e}$	$c_{1}$	$c_{2}$	$r_{f}$	$a_{d}$	$r_{h}$	$d_{h}$	$t_{h}$	$t_{s}$	$t_{b}$	$no_{b}$	$n o$
Line	$x$	$y$
Arc	$x$	$y$	$\alpha$	$f$
Circle	$x$	$y$			$r$
Spline
SCP	$x$	$y$
Extrude						$\theta$	$\phi$	$\gamma$	$p_{x}$	$p_{y}$	$p_{z}$	$s$	$e_{1}$	$e_{2}$	$u_{1}$	$u_{2}$			$b$
Topo
Selection																													$t y$	$b\_t$	$b\_n$	$n o$
Revolve						$\theta$	$\phi$	$\gamma$	$p_{x}$	$p_{y}$	$p_{z}$	$s$					$e_{1}$	$e_{2}$	$b$
Pocket						$\theta$	$\phi$	$\gamma$	$p_{x}$	$p_{y}$	$p_{z}$	$s$	$e_{1}$	$e_{2}$	$u_{1}$	$u_{2}$
Groove						$\theta$	$\phi$	$\gamma$	$p_{x}$	$p_{y}$	$p_{z}$	$s$					$e_{1}$	$e_{2}$
Shell																				$t_{1}$	$t_{2}$
Chamfer																						$e_{1}$	$e_{2}$
Fillet																								$r$
Hole									$x$	$y$	$z$															$r$	$d e$	$t y$
Mirror
Mirror topo
Draft																									$a n$

The definition of $p_{i}$ is as follows:

\displaystyle p_{i}=(x,y,\alpha,f,r,\theta,\phi,\gamma,p_{x},p_{y},p_{z},s,e_{% 1},e_{2},et_{1},et_{2},a_{1},a_{2},b,t_{i},t_{e},c_{1},c_{2},r_{f},a_{d},r_{h}% ,d_{h},t_{h},t_{s},t_{b},no_{b},no)

(5)

which describes the parametric features of the particular $t_{i}$ .

The corresponding relationship between the feature vector and the definition of features is shown in Table 2. When designing the number of bits occupied by parameters with different features, we try not to reuse them as much as possible. However, when containing the same meaning such as coordinates or angles, a small amount of bit reuse is allowed. In addition, due to the special characteristic of the selection feature, we allocate 4 exclusive bits for it.

3.3 Simple features, advanced features (selection feature and advanced features depend on selection), and the corresponding parameters

Table 1 describes the simple features, advanced features (selection feature and advanced features depend on selection), and their corresponding parameters established by WHUCAD dataset. There are 4 times more feature types in the WHUCAD dataset than in the DeepCAD dataset and the Fusion 360 Gallery dataset. The feature operations and parameters of WHUCAD are as complicated as those in industry CAD systems. Thus, the most important value of the WHUCAD dataset is not the data size but the fruitful feature types for supporting HCI.

Line. The line feature has two parameters, $x$ and $y$ , which represent the coordinates of the endpoint.

Arc. The parameters of the arc feature include the coordinates $x$ and $y$ of the arc endpoint, the sweep angle $\alpha$ , and the clockwise and counterclockwise markers $f$ (1 represents counterclockwise).

Circle. The parameters of the circle feature include the center coordinates $x$ , $y$ , and radius $r$ .

Spline. The spline feature declares the Spline operation, followed by the declaration of the spline control point (SCP), which contains the coordinates $x$ and $y$ of each spline point except the starting point.

Extrude. Extrude is conducted on a sketch to create a 3D CAD model once the sketch is completed, where $p_{x}$ , $p_{y}$ , and $p_{z}$ are the absolute coordinates of the sketch origin, $\theta$ , $\phi$ , and $\gamma$ are the normal vector directions of the selected sketch, $s$ is the size of the sketch, $e_{1}$ and $e_{2}$ are the lengths of the stretch in both directions, $b$ represents the Boolean operation, $u_{1}$ and $u_{2}$ show types of stretching in both directions (includes blind holes, stretching to the next face, stretching to a specified face, etc.).

Selection. The selection feature is one of the most important features in WHUCAD. All feature operations except for extrude and selection itself depend on the selection feature. For example, a fillet operation or a revolve operation needs to be conducted on a specific object, such as a line or a face. In this case, the selection operation is used to specify which part will be chosen for the next operation. A selection feature consists of four parameters: $t y p e$ , $entity\_type$ , $entity\_no$ and $n o$ , where $t y p e$ indicates the type of the selected object (wire, face, edge, etc.), $entity\_type$ indicates the type of the geometry in which the object is located (sketch, extrude, revolve, shell, chamfer, etc.), $entity\_no$ indicates the serial number of the geometry, and $n o$ indicates the serial number of the selected object itself.

Revolve feature, Pocket feature and Groove feature have the same parameter meanings as the extrude feature.

Topo. This feature marks the beginning of a selection object declaration, followed by a series of selection instructions.

Shell. The parameters of the shell feature include $t_{1}$ and $t_{2}$ , which indicate the inner and outer thicknesses of the extraction shell, respectively, and are usually preceded by the selection command.

Chamfer. The chamfer parameters include $e_{1}$ and $e_{2}$ , which indicate the two chamfer lengths, respectively, and are generally preceded by the selection command.

Fillet. The feature parameter $r$ indicates the radius of the rounded corner, which is usually preceded by the selection command.

Hole. This feature has six parameters, where $x$ , $y$ , and $z$ represent the absolute position of the center. $r$ is the radius of the hole, $d e p t h$ indicates the depth of the hole feature, and $t y p e$ is the type of hole (for example blind holes, until the next face, until the specified face, etc.).

Mirror. This feature is usually preceded by a selection command, which indicates to flip the object along a specified axis of symmetry.

Mirror Topo. This feature is used before the selection command to indicate that the selected object is a mirror object.

Draft. This feature has a parameter $a n g l e$ , which represents the angle of pulling the mold, which follows the selection command. The first object is the neutral plane, and the remainder are the target planes. The default reference direction vector is the normal vector of the neutral plane.

Furthermore, the proposed WHUCAD dataset is scalable because of the development of mainstream CAD systems.

3.4 Selection feature to learn and imitate HCI process under deep learning network

The WHUCAD dataset contains a selection feature that can learn and imitate the HCI process to focus on and select a particular topological entity for the next feature operation. In this way, the design intention and design knowledge of a CAD engineer can be fully learned in a deep learning network.

Figure 7.

An example of the mirror feature.

The selection mechanism is important because it establishes a relationship between two topological elements. The existing learning-based 3D modeling tasks, such as reconstruction and generation, can only reach the solid modeling stage and cannot reach the parametric modeling stage. For example, as shown in Fig. 8, a small cuboid is created in the center of the top surface of the large cuboid. Due to the lack of a selection mechanism, in DeepCAD, once the extrusion parameter of the large cuboid is decreased, two related entities will be separated. The CAD designer must manually move the top cuboid downward and reposition it on the top surface of the large cuboid. However, a true parameterized modeling process should allow the small cuboid to move as the extrusion parameters of the large cuboid decrease. This situation will be solved in the WHUCAD. Figure 9 shows this situation in WHUCAD. Due to the naming mechanism, the top cuboid and the bottom cuboid are truly linked together through the parameterized modeling process.

Figure 8.

When modifying parameters in DeepCAD data, as the two extruded sketches record absolute coordinate information, reducing one of the extruding parameters can cause the two bodies to separate.

Figure 9.

In WHUCAD, the sketch of the second extrusion is created by selecting the plane of the first extrusion. This selection operation associates the coordinate information of the second sketch with that of the previous entity. Reducing the parameter of the previous extrusion does not result in the separation of the two entities. This is what a true parametric modeling should achieve, and aligns with the design intent in most scenarios.

The proposed selection and name mechanism can be input into a deep learning network to learn how a real CAD engineer selects an entity in CAD software. The selection feature and the name mechanism are not available in any existing 3D datasets, i.e., they are available only in the WHUCAD dataset.

Typical examples of selection and the name mechanism in parametric and feature-based modeling history are illustrated in the following.

(1) How the selection and name mechanism work when extruding a sketch into a 3D model

As shown in Fig. 10(a), regardless of how a rectangle is sketched, we name every edge in a counterclockwise direction.

According to the feature definitions in Table 1, the selection feature operation has four parameters, which are the type of object, the type of entity, the serial number of the entity, and the serial number of the selected object, respectively. For the case of Fig. 10(a), the four edges can be named “sketch.1;1”, “sketch.1;2”, “sketch.1;3”, and “sketch.1;4”. For further illustration of “sketch.1;1”, the detailed name structure is shown in Table 3.

After selecting the sketch, an extrude feature operation is executed to create a number of faces, which construct a 3D box.

As shown in Fig. 10(b), once the box is constructed, these new faces are also named and can be selected for further feature operation.

Figure 10.

An example of the selection feature and naming mechanism for selecting edges from the sketch and new faces from extrude.

(2) How the selection and name mechanism work when conducting a chamfer operation on an edge of the box.

As shown in Fig. 11, the edge between the face of extrude.1;2 and extrude.1;0:(sketch.1;1) can be named “(extrude.1;2);(extrude.1;0 $:$ (sketch.1;1))”. Therefore, this edge can be selected for feature operation.

After a chamfer operation is executed on the edge “(extrude.1;2);(extrude.1;0 $:$ (sketch.1;1))”, a new face is created and named as “chamfer.1;((ext rude.1;2); (extrude.1;0: (sketch.1;1)))”. The detailed name structure of the newly created chamfer face is shown in Table 4. Using the name, the new chamfer face can be selected for the subsequent operation.

Figure 11.

An example of the selection feature and naming mechanism for selecting a chamfer.

Table 3

Command instruction of sketch

Commands	Selection type	Entity type	Entity no	No
Selection	Wire	Sketch	1	1

Table 4

Command instruction of chamfer

Commands	Selection type	Entity type	Entity no	No
Selection	Face	Extrude	1	2
Selection	Wire	Sketch	1	1
Selection	Face	Extrude	1	0
Selection	Face	Chamfer	1	0

(3) How the selection and name mechanism work when a face is split by a previous operation.

In Fig. 12, a semicylinder (Fig. 12(b)) is extruded from a closed semicircle (Fig. 12(a)), and then a pocket operation is conducted to another closed semicircle with a smaller radius (Fig. 12(c)). Thus, face F1 is divided into two parts. F1_1 and F1_2 are both from the original face F_1 and the pocket operation (Although some other methods regard the pocket operation as the same result as conducting an extrude operation with Boolean abstraction; thus, they abandon the pocket as a feature. However, the design intention and design knowledge of pocket features are quite different from extrude. Furthermore, if the WHUCAD dataset is applied to classification or segmentation tasks, there will be a significant difference in real meaning between pocket and extrude.) From the basic cases in the previous part, the names of W1 and W2 are “sketch.1;1” and “sketch.1;2” respectively. Here we denote W1 $=$ Wire(sketch.1;1) and W2 $=$ Wire(sketch.1;2) (“Wire” is the selection type when a selection command is conducted on the edge of a sketch. Another selection type “Edge” is used to select an edge of a 3D CAD model.) The name of F1 is “extrude.1;0:(sketch.1;1)”. For readability, we denote F1 $=$ Face(extrude.1;0:W1) (“Face” also indicates the selection type). Thus, we denote the created face F5 $=$ Face(pocket.1;0:Wire(sketch.2;2)). (We did not display the second sketch of semicircle with a smaller radius, the arc part of which is named as “sketch.2;2”.) In this way, F1_1 is denoted as “Face(F1;[F5;F4;F2;F3])”. The order of “[F5;F4;F2;F3]” is the counterclockwise direction of faces adjacent to F1_1 (starting from the newly created face F5 by pocket operation). F1_2 is denoted as “Face(F1;[F5;F3;F2;F4])”.

Figure 12.

How the selection and name mechanism work when a face is split by a previous operation.

3.5 Creation of the WHUCAD dataset

Although DeepCAD and Fusion 360 Gallery consider CAD processing as command sequences, they have some critical problems. While DeepCAD claims to have 178,238 data, after removing redundant models (just simply removing the same models by comparing each command sequence), there remains 131,553 models remain. These 46,685 redundant data are not conducive to training and may cause overfitting. In addition, the models in DeepCAD only contain a few simple features and several steps. The complexities of the models are generally low. The Fusion 360 Gallery dataset has only 8,625 CAD models, which is insufficient for well training a deep learning network.

WHUCAD consists of two subsets. The first subset is derived from DeepCAD. To address the simplicity of DeepCAD, we augment the data by randomly inserting sequences containing advanced features into DeepCAD data. The advantage of this approch is that it can obtain a rich amount of data with advanced features for learning and complex processing steps. We filter out unreasonable augmented data by integrating with industrial CAD software such as CATIA to ensure that every newly created model with advanced feature types can be visualized and converted to B-rep, mesh, and point cloud data. Furthermore, we check every model to ensure that WHUCAD has no redundant data. The second subset comes from manual construction. Our team spent more than two years building 4,244 CAD models from the ABC dataset that have fruitful advanced features and semantic meaning but cannot be parsed into a command sequence by DeepCAD due to the lack of a selection mechanism. All models in the second subset of WHUCAD are macro recorded by CATIA and then parsed into vector sequence format. These two subsets constitute WHUCAD, and each has its own characteristics. We annotate the two subsets in the final merged WHUCAD dataset and will open-source all code at Github.

We provide WHUCAD data in .h5 format. Each file represents a CAD model. Each row in the CAD file represents a feature operation. The relationships between each digit and its meaning are shown in Table 2. These data in .h5 files can be directly as input data to deep learning network for training and testing. Furthermore, we provide a parsing tool along with the source code to integrate WHUCAD with major industrial software such as CATIA for visualization and further edit for researchers.

We will also issue a license to researchers who require WHUCAD for academic research.

3.6 Discussion of feature types, history length and type distribution

Compared with DeepCAD, WHUCAD is much richer in terms of types of features and has a more reasonable proportional distribution. A distribution of complicated feature types selected from our 146,066 models is shown in Fig. 13.

Figure 13.

Distribution of complicated feature types in 4,244 manually constructed subset from WHUCAD.

Figure 14.

Comparison of the modeling history complexity of individual CAD models (creation steps) between DeepCAD (left) and WHUCAD (right).

The length of the modeling history is an important metric for evaluating the complexity of a CAD model. Figure 14 shows the history length of each CAD model between DeepCAD and WHUCAD, in which the WHUCAD models have a longer history than DeepCAD. Therefore, WHUCAD models are more complicated and contain more features.

Figure 15.

Distribution of complicated features between the WHUCAD and DeepCAD datasets.

In addition, the distribution of feature types in the WHUCAD dataset complies with the thinking and reasoning of human engineers. The extrude feature is the most frequently used feature. Features that human engineers use a lot are fillet and chamfer. Therefore, the overall distribution of WHUCAD feature types is more reasonable than that of DeepCAD as shown in Fig. 15.

4. Experiment

To the best of our knowledge, at present, only the DeepCAD and Fusion 360 Gallery reconstruction datasets are historical 3D datasets in which the CAD models are described as a sequence of CAD commands.

Therefore, we compare the WHUCAD dataset with the DeepCAD and Fusion 360 Gallery datasets in the reconstruction and generation tasks. In Section 4.1, we present quantitative experiments of the reconstruction task. In Section 6.1, we present qualitative experiments of the generation task. The network architecture of the reconstruction and generation tasks is shown in Fig. 16. In Section 6.2, we present industrial experiments on the WHUCAD dataset with the modern interactive CAD software.

Figure 16.

The network architecture of the reconstruction and generation experiments.

4.1 Quantitative experiments

In the quantitative experiments, we train the WHUCAD, DeepCAD, and Fusion 360 Gallery reconstruction datasets in the original DeepCAD network and the WHUCAD dataset in our proposed network, as shown in Fig. 16. The process in the orange dashed box is the reconstruction pipeline. The command sequences are first parsed into advanced feature sequences (the vector format) using a parsing python script, which will be open source with the training and testing code. The vector sequences are not only transmitted to the embedding process but also used as input data to the generation pipeline. After the transformer encoder, the latent $z$ contains the latent features of the advanced feature sequences. Then, $z$ will be put into a transformer decoder. Through linear layers, the predicted feature sequences are obtained.

The equation of the loss function between the predicted sequences and ground truth is defined as follows:

\displaystyle L=L_{\textit{cmd}}+\beta L_{\textit{param}}

(6)

where:

L_{𝑐𝑚𝑑} = \sum_{i = 1}^{N_{c}} l [{\hat{t}}_{i}, t_{i}]

(7)

L_{𝑝𝑎𝑟𝑎𝑚} = \sum_{i = 1}^{N_{c}} \sum_{j = 1}^{N_{p}} l [{\hat{p}}_{i, j}, p_{i, j}]

(8)

and $N\_c$ is the maximum number of CAD commands (100 in WHUCAD), and $N_{p}$ is the number of parameters (32 in WHUCAD), $l$ is the standard Cross-Entropy [74], and $\beta$ is a hyper parameter.

Table 5

Results of measuring the quality of reconstruction

Evaluation	Ours with stricter limit*	Ours*	Ours†	DeepCAD	Fusion 360 gallery
$\textit{ACCcmd}\uparrow$	99.45%	99.45%	99.25%	99.34%	88.83%
$\textit{ACCparam}\uparrow$	98.17%	98.22%	55.08%	97.47%	81.23%
$\textit{median CD}(*1000)\downarrow$	0.137	0.137	0.220	0.193	1.159

$\uparrow$ : the higher is better. $\downarrow$ : the lower is better. * indicates WHUCAD is trained on the network shown in Fig. 16 with capability to learn advanced features. †indicates WHUCAD is trained on the original DeepCAD network which can only learn simple features.

Figure 17.

Generated CAD models from the WHUCAD dataset with feature notations.

Figure 18.

Generated CAD models from the DeepCAD dataset.

In the network architecture, the number of heads in multihead attention is set to 8. The dimensions of the embedded feature sequence and latent vector z are 256. The training process adopts the Adam optimizer with a learning rate of 0.001. The dropout rate is set to 0.1.

Similar to DeepCAD, both WHUCAD and Fusion 360 Gallery are split into training, validation, and test sets by 90%-5%-5%. We train the three datasets for 1,000 epochs with a batch size of 512. All the experiments are conducted on a PC with an Intel I9-12700 CPU and an NVIDIA GeForce RTX 3090ti GPU. The quantitative results of the reconstruction experiments are shown in Table 5.

5. Metrics of ACCcmd and ACCparam

ACCcmd and ACCparam represent the correctness of the predicted CAD command (feature operation) type and the correctness of the command (feature operation) parameters when the command type is correctly predicted.

ACCcmd is defined as:

\displaystyle\textit{ACC}_{\textit{cmd}}=\frac{1}{N_{c}}\sum\limits_{i=1}^{N_{% c}}\mathcal{I}[\hat{c}_{i},c_{i}]

(9)

where $N_{c}$ denotes the total number of CAD commands, $c_{i}$ is the predicted command, $\hat{c}_{i}$ is the ground truth, and $\mathcal{I}$ is the indicator function, which takes a value of 0 or 1.

Once the command type is predicted correctly, the command parameter is calculated. The definition of ACCcmd is as follows:

\displaystyle\textit{ACC}_{\textit{param}}=\frac{1}{K}\sum\limits_{i=1}^{N_{c}% }\sum\limits_{j=1}^{N_{p}}\mathcal{I}[\left|\hat{p}_{i,j}-p_{i,j}\right|<\eta]% \mathcal{I}[c_{i}=\hat{c}_{i}]

(10)

where $K$ is the total number of parameters in the corrected predicted commands, $\hat{p}_{i,j}$ is the ground truth of each parameter, $p_{i,j}$ is the predicted parameter, and $\eta$ is a tolerance threshold of the error value.

For a fair comparison, we choose to remain consistent with DeepCAD, which sets $\eta=$ 3. In Table 5, “Ours†” indicates that WHUCAD is trained on the original DeepCAD network, which has the ability to learn only simple feature parameters. It is clear from the ACCparam results that WHUCAD is more difficult to learn due to its fruitful advanced features. This result further demonstrates that WHUCAD has greater value than DeepCAD in learning complex and advanced features and can also be applied to advanced CAD feature learning tasks.

To solve the problem of DeepCAD being unable to learn advanced features, in the “Ours*” experiment shown in Table 5, we incorporate the ability to process advanced features when converting “complicated command sequences” to “advanced feature sequences”, which was demonstrated in Section 3. The experimental results show that we achieved better results than DeepCAD and Fusion 360 Gallery by using our networks with advanced feature learning abilities.

Figure 19.

Generated CAD models from the Fusion 360 Gallery dataset.

Figure 20.

The process of editing details of generated models from the WHUCAD dataset.

However, to more rigorously verify the effectiveness of feature learning, we consider that a small fluctuation of the parameter in the selection operation may cause the wrong edge or face to be selected. Therefore, for a stricter requirement, we add an extra experiment that does not allow any errors in predicting the parameters of the selection feature. In Table 5, “Ours with stricter limit*” indicates that the predicted parameters in the selection operation do not allow any errors from the ground truth. The results show that even under the requirement of allowing zero errors in the parameters of the selection feature, the results of ACCparam on the WHUCAD dataset trained by our network are still better than those on the DeepCAD and Fusion 360 Gallery datasets.

The prediction rates of ACCcmd and ACCparam are both important. Only when their accuracy is simultaneously high can it be concluded that the model’s predictions are relatively accurate. For example, for ACCcmd, when a rectangular sketch is selected and an extrusion operation is desired, if the intended “extrude” feature type is mistakenly predicted as “revolve”, it will lead to an overall error in the model. This is because the bits of parameters required for the revolve feature differ significantly from those of the extrude feature (the specific feature types and the corresponding bits of parameter are shown in Table 2). Therefore, ACCcmd is crucially important. However, if a topological element is supposed to be revolved by a small angle and the revolve type is correctly predicted, but the parameter is incorrectly predicted for a very large angle, the resulting model will also be significantly different. This indicates that the ACCparam prediction is equally important. Therefore, the ACCcmd and ACCparam are crucial metrics for evaluating whether a model has been accurately constructed. Both are important and indispensable.

6. Metric of chamfer distance

From another perspective to measure the quality of the recovered 3D geometry, the chamfer distance (CD) is chosen for evaluation by uniformly sampling 8000 points on the surfaces of the reference shape and the recovered shape. The equation of the chamfer distance is as follows:

\displaystyle\textit{CD}(S_{1},S_{2})=\frac{1}{\left|S_{1}\right|}\sum_{x\in S% _{1}}\min_{y\in S_{2}}\left\|x-y\right\|_{2}^{2}

(11)

where $S_{1}$ and $S_{2}$ denote the point cloud of the reconstructed CAD model and the point cloud of the ground truth model, respectively, and $|S_{1}|$ and $|S_{2}|$ represent the number of the points.

When using the DeepCAD original network to train the WHUCAD dataset (“Ours†”), WHUCAD has a higher median CD, which also indicates that models in WHUCAD are generally more complex and harder to reconstruct than those in DeepCAD. In addition, conducting experiments under our network (Ours*) has a lower median CD than DeepCAD, which proves the superiority of our network.

In conclusion, in Table 5, for quantitative comparisons, experiments “Ours†”, “DeepCAD”, and “Fusion 360 Gallery” indicate that models in WHUCAD are generally more complicated and harder to learn than DeepCAD and Fusion 360 Gallery, which reflects the value of WHUCAD. Advanced features and the key selection mechanism in WHUCAD provide additional possibilities and new benchmarks for parametric and feature-based CAD tasks. Furthermore, comparisons among the experiments “Ours*”, “Ours†”, and “DeepCAD” indicate the effectiveness of our network for advanced feature learning.

6.1 Qualitative experiments

Once the reconstruction pipeline is well trained, the weight parameters in the decoder are fixed. To generate CAD models, advanced feature sequences in vector format are input to the generation pipeline (the process is shown in the green dashed box) for training. Similar to DeepCAD, we use the Wasserstein GAN network [23, 75, 76] to train the input feature sequences. The generator and discriminator of the GAN are both MLP networks with four hidden layers. After sufficient training, random vectors sampled from a multivariate Gaussian distribution are fed into the GAN generator to generate the latent vector $z$ . Then, the generated $z$ is input into the trained transformer decoder and the linear layers to generate new feature sequences. We also provide related parsing tools to support visualization and editing in mainstream CAD software (CATIA, SolidWorks, and Fusion 360). In this way, varieties of new parametric and feature-based historical CAD models are created.

Figure 17 shows examples of CAD models generated from WHUCAD. These examples include real-world engineering features, such as fillet, chamfer, shell, pocket, draft, mirror, revolve, hole, and groove, which are as advanced as those in CAD models created by a real human engineer who interactively works on modern industrial CAD software. For better visualization, notations for advanced features are added to the above examples.

The models generated from the DeepCAD and Fusion 360 Gallery datasets are shown in Figs 18 and 19, respectively.

According to Figs 17 to 19, for qualitative comparisons, it is clearly that the 3D shapes generated from WHUCAD are more complicated, more advanced, more vivid, and more detailed than those from the DeepCAD and Fusion 360 Gallery datasets.

6.2 Industrial experiments

The WHUCAD dataset is compatible and transferable with modern interactive CAD software.

The CAD models generated from the WHUCAD dataset can be easily imported into industrial software. Then, the input CAD models can be edited and redesigned by a human engineer who works on historical and feature-based CAD software, such as CATIA,3

³
https://www.3ds.com/products-services/catia/.
SolidWorks,4 ⁴
https://www.solidworks.com/.
and Auto-desk Fusion 360.5 ⁵
https://www.autodesk.com/.

Figure 20 shows an HCI editing process of an input CAD model generated from WHUCAD on one industrial CAD software.

Step 1: A human engineer imports an initial CAD model (a phone protection case) generated from the WHUCAD dataset into the software.

Step 2: To adjust the size of the camera frame, the human engineer switches the 3D model to the related sketch plane.

Step 3 to step 9: The human engineer selects the edges of the camera frame one by one. After selecting each edge, the human engineer modifies the length of the sides and the center coordinates of the fillet operations.

Thus, WHUCAD has real-world engineering ability and can support collaborative product development (CPD) in modern globalization era.
7. Conclusion

In this paper, we provide WHUCAD, a full parametric and historical feature-based CAD dataset with fruitful advanced features, which enables deep learning to learn HCI in 3D CAD learning. In contrast to existing datasets with simple CAD commands (line, arc, circle, and extrude), WHUCAD includes many more rich advanced feature types (such as selection, revolve, pocket, groove, shell, chamfer, fillet, hole, mirror, and draft). More importantly, the existing CAD datasets and methods only support overall extrusion on sketches, without the characteristics of HCI. Therefore, deep learning networks are also unable to learn and mimic HCIs to select specific topological elements for more advanced features. To compensate for this deficiency, WHUCAD first devises a selection feature and a name mechanism that can be input into a deep learning network to learn how a human CAD engineer selects an entity in CAD software. Quantitative experiments, qualitative experiments, and industrial experiments verify the superiority of WHUCAD. Furthermore, we first provide a parsing tool with a parametric and feature-based CAD dataset for integrating our parametric CAD models with industrial software, which can help human engineers truly implement edits and redesigns of historical and parametric models.

To the best of our knowledge, WHUCAD is the first parametric and feature-based CAD dataset to support HCI in 3D learning. In the future, we will continue to study the limitations commonly found in existing CAD models. For example, the feature types have not fully covered all types in the real world, and there is still a lack of parameterized feature-based datasets for classification. We will continue to extend the WHUCAD dataset to support more complicated HCI in 3D learning. Additionally, we will combine WHUCAD with advanced machine learning methods [77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87] for knowledge-based CAD/CAM engineering applications [88, 89, 59].

Footnotes

Acknowledgments

This work is supported by the National Natural Science Foundation of China under Grant No. 62072348, and the Shenzhen Polytechnic University Research Fund under Grants 6023310030K, 6022312044K, 6023240118K, and 6024310045K. The numerical calculations in this paper were performed on the supercomputing system at the Supercomputing Center of Wuhan University.

References

Kim

Moon

Lee

Mun

. End-to-end digitization of image format piping and instrumentation diagrams at an industrially applicable level. Journal of Computational Design and Engineering.2022; 9(4): 1298-1326.

Cheng

Cuvin

Olechowski

Zhou

. User Perspectives on Branching in Computer-Aided Design. Proceedings of the ACM on Human-Computer Interaction.2023; 7(CSCW2): 1-30.

García Gómez

González-Rodríguez

Vela

. Enhanced memetic search for reducing energy consumption in fuzzy flexible job shops. Integrated Computer-Aided Engineering.2023; 30(2): 151-167.

Martins

Papa

Adeli

. Deep learning techniques for recommender systems based on collaborative filtering. Expert Systems.2020; 37(6): e12647.

Rosso

Aloisio

Randazzo

Tanzi

Cirrincione

Marano

. Comparative deep learning studies for indirect tunnel monitoring with and without Fourier pre-processing. Integrated Computer-Aided Engineering.2024; 31(2): 213-232.

Fernández-Rodríguez

Palomo

Ortiz-de Lazcano-Lobato

Ramos-Jiménez

López-Rubio

. Dynamic learning rates for continual unsupervised learning. Integrated Computer-Aided Engineering.2023; 30(3): 257-273.

Colonnese

Di Luzio

Rosato

Panella

. Bimodal Feature Analysis with Deep Learning for Autism Spectrum Disorder Detection. International Journal of Neural Systems.2023; pp. 2450005-2450005.

Villarrubia-Martin

Rodriguez-Benitez

Jimenez-Linares

Muñoz-Valero

Liu

. A hybrid online off-policy reinforcement learning agent framework supported by transformers. International Journal of Neural Systems.2023; 33(12): 2350065.

Albera

Le Bouquin Jeannes

Kachenoura

Karfoul

Yang

, et al. Epileptic seizure prediction using deep neural networks via transfer learning and multi-feature fusion. International Journal of Neural Systems.2022; 32(07): 2250032.

10.

Xie

Zhang

. Deep multimodal neural network based on data-feature fusion for patient-specific quality assurance. International journal of neural systems.2022; 32(01): 2150055.

11.

Hassanpour

Moradikia

Adeli

Khayami

Shamsinejadbabaki

. A novel end-to-end deep learning scheme for classifying multi-class motor imagery electroencephalography signals. Expert Systems.2019; 36(6): e12494.

12.

Nogay

Adeli

. Machine learning (ML) for the diagnosis of autism spectrum disorder (ASD) using brain imaging. Reviews in the Neurosciences.2020; 31(8): 825-841.

13.

Jones

Kodnongbua

Kim

Schulz

. Self-supervised representation learning for cad. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023; pp. 21327-21336.

14.

Wan

Song

Schonfeld

. A geographic information model for 3-D environmental suitability analysis in railway alignment optimization. Integrated Computer-Aided Engineering.2023; 30(1): 67-88.

15.

Qin

Gao

Yang

Chen

. A deep learning approach to the classification of 3D CAD models. Journal of Zhejiang University SCIENCE C.2014; 15(2): 91-106.

16.

Guo

Liu

Pan

Liu

Tong

Guo

. Complexgen: Cad reconstruction by b-rep chain complex generation. ACM Transactions on Graphics (TOG).2022; 41(4): 1-18.

17.

Xie

Hao

Zhao

Liang

. A spatio-temporal method for extracting gamma-band features to enhance classification in a rapid serial visual presentation task. International Journal of Neural Systems.2022; 32(03): 2250010.

18.

Zhang

Rong

Paul

Zhu

Neri

, et al. A layered spiking neural system for classification problems. International Journal of Neural Systems.2022; 32(08): 2250023.

19.

Rafiei

Adeli

. A new neural dynamic classification algorithm. IEEE Transactions on Neural Networks and Learning Systems.2017; 28(12): 3074-3083.

20.

Pereira

Piteri

Souza

Papa

Adeli

. FEMa: A finite element machine for fast learning. Neural Computing and Applications.2020; 32: 6393-6404.

21.

Alam

KMR

Siddique

Adeli

. A dynamic ensemble learning algorithm for neural networks. Neural Computing and Applications.2020; 32(12): 8675-8690.

22.

Rafiei

Gauthier

Adeli

Takabi

. Self-supervised learning for electroencephalography. IEEE Transactions on Neural Networks and Learning Systems.2022.

23.

Xiao

Zheng

. Deepcad: A deep generative network for computer-aided design models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision; 2021; pp. 6772-6782.

24.

Willis

Luo

Chu

Lambourne

, et al. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM Transactions on Graphics (TOG).2021; 40(4): 1-24.

25.

Urdiales

Martín

Armingol

. An improved deep learning architecture for multi-object tracking systems. Integrated Computer-Aided Engineering.2023; 30(2): 121-134.

26.

Daranda

Dzemyda

. Reinforcement learning strategies for vessel navigation. Integrated Computer-Aided Engineering.2023; 30(1): 53-66.

27.

Chang

Funkhouser

Guibas

Hanrahan

Huang

, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv151203012. 2015.

28.

Geometry

. Neural Shape Parsers for Constructive Solid Geometry. Ieee Transactions on Pattern Analysis and Machine Intelligence.2022; 44(5).

29.

Sharma

Goyal

Liu

Kalogerakis

Maji

. Csgnet: Neural shape parser for constructive solid geometry. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; 2018; pp. 5515-5523.

30.

Inala

Spielberg

Schulz

Rus

, et al. Inversecsg: Automatic conversion of 3d models to csg trees. ACM Transactions on Graphics (TOG).2018; 37(6): 1-16.

31.

Guibas

. Pointnet: Deep learning on point sets for 3d classification and segmentation. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017, pp. 652-660.

32.

Guibas

. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in Neural Information Processing Systems.2017; 30.

33.

Mohammadi

Wang

Del Bue

. Pointview-gcn: 3d shape classification with multi-view point clouds. In: 2021 IEEE International Conference on Image Processing (ICIP). IEEE; 2021, pp. 3103-3107.

34.

Song

Duan

Liang

Yan

. A kernel correlation-based approach to adaptively acquire local features for learning 3D point clouds. Computer-Aided Design.2022; 146: 103196.

35.

Kania

Zieba

Kajdanowicz

. UCSG-NET-unsupervised discovering of constructive solid geometry tree. Advances in Neural Information Processing Systems.2020; 33: 8776-8786.

36.

Chen

Sanghi

Shayani

Mahdavi-Amiri

, et al. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022, pp. 11768-11778.

37.

Zhou

Jacobson

. Thingi10k: A dataset of 10,000 3d-printing models. arXiv preprint arXiv160504797. 2016.

38.

Bogo

Romero

Pons-Moll

Black

. Dynamic FAUST: Registering human bodies in motion. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2017; pp. 6233-6242.

39.

Song

Duan

Bai

. LSLPCT: An Enhanced Local Semantic Learning Transformer for 3-D Point Cloud Analysis. IEEE Transactions on Geoscience and Remote Sensing.2022; 60: 1-13.

40.

Song

Khosla

Zhang

Tang

, et al. 3d shapenets: A deep representation for volumetric shapes. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2015; pp. 1912-1920.

41.

Lahav

Tal

. Meshwalker: Deep mesh understanding by random walks. ACM Transactions on Graphics (TOG).2020; 39(6): 1-13.

42.

Afham

Dissanayake

Dharmasiri

Thilakarathna

Rodrigo

. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022; pp. 9902-9912.

43.

Koch

Matveev

Jiang

Williams

Artemov

Burnaev

, et al. Abc: A big cad model dataset for geometric deep learning. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition; 2019; pp. 9601-9611.

44.

Wang

Tagliasacchi

Zhou

Mahdavi-Amiri

, et al. Pie-net: Parametric inference of point cloud edges. Advances in Neural Information Processing Systems.2020; 33: 20167-20178.

45.

Ganin

Bartunov

Keller

Saliceti

. Computer-aided design as language. Advances in Neural Information Processing Systems.2021; 34: 5885-5897.

46.

Wen

Zhang

. Pixel2mesh++: Multi-view 3d mesh generation via deformation. In: Proceedings of the IEEE/CVF international conference on computer vision; 2019; pp. 1042-1051.

47.

Boulch

Marlet

. Poco: Point convolution for surface reconstruction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022; pp. 6302-6314.

48.

Bazazian

Parés

. EDC-Net: Edge detection capsule network for 3D point clouds. Applied Sciences.2021; 11(4): 1833.

49.

Liu

Yang

Guo

. Surface and Edge Detection for Primitive Fitting of Point Clouds. In: ACM SIGGRAPH 2023 Conference Proceedings. SIGGRAPH ’23. Association for Computing Machinery; 2023. Available from: doi: 10.1145/3588432.3591522.

50.

Sanghi

. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16. Springer; 2020. pp. 626-642.

51.

Kalogerakis

Simari

Nowrouzezahrai

Singh

. Robust statistical estimation of curvature on discretized surfaces. In: Symposium on Geometry Processing. vol. 13; 2007, pp. 110-114.

52.

Boulch

Marlet

. Fast and robust normal estimation for point clouds with sharp features. In: Computer graphics forum. vol. 31; Wiley Online Library; 2012, pp. 1765-1774.

53.

Kostrikov

Jiang

Panozzo

Zorin

Bruna

. Surface networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition; 2018, pp. 2540-2548.

54.

Tessier

Wang

. Ontology-based feature mapping and verification between CAD systems. Advanced Engineering Informatics.2013; 27(1): 76-92.

55.

Maquart

Elguedj

Gravouil

Rochette

. 3D B-Rep meshing for real-time data-based geometric parametric analysis. Advanced Modeling and Simulation in Engineering Sciences.2021; 8: 1-28.

56.

Jayaraman

Lambourne

Desai

Willis

Sanghi

Morris

. Solidgen: An autoregressive model for direct b-rep synthesis. arXiv preprint arXiv220313944. 2022.

57.

Zhu

Menq

. B-rep model simplification by automatic fillet/round suppressing for efficient automatic feature recognition. Computer-Aided Design.2002; 34(2): 109-123.

58.

Bianconi

Conti

Di Angelo

. Interoperability among CAD/CAM/CAE systems: a review of current research trends. Geometric Modeling and Imaging–New Trends (GMAI’06). 2006; pp. 82-89.

59.

Zhang

. Service-Oriented Feature-Based Data Exchange for Cloud-Based Design and Manufacturing. IEEE Transactions on Services Computing.2018; 11(2): 341-353.

60.

Lambourne

Willis

Jayaraman

Zhang

Sanghi

Malekshan

. Reconstructing editable prismatic cad from rounded voxel models. In: SIGGRAPH Asia 2022 Conference Papers; 2022. pp. 1-9.

61.

Zhang

Guan

Jiang

Ning

Wang

Tan

. Brep2Seq: a dataset and hierarchical deep learning network for reconstruction and generation of computer-aided design models. Journal of Computational Design and Engineering.2024; 11(1): 110-134.

62.

Zhou

. MultiCAD: Contrastive Representation Learning for Multi-modal 3D Computer-Aided Design Models. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. CIKM ’23. Association for Computing Machinery; 2023. pp. 1766-1776. Available from: doi: 10.1145/3583780.3614982.

63.

Willis

Lambourne

Cheng

Jayaraman

Furukawa

. SkexGen: Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks. In: International Conference on Machine Learning. PMLR;2022; pp. 24698-24724.

64.

Lambourne

Willis

KDD

Jayaraman

Sanghi

Meltzer

Shayani

. BRepNet: A Topological Message Passing System for Solid Models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2021; pp. 12773-12782.

65.

Willis

Jayaraman

Chu

Tian

Grandi

, et al. JoinABLe: Learning Bottom-up Assembly of Parametric CAD Joints. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society; 2022, pp. 15828-15839.

66.

Guo

Zhang

Yan

. Secad-net: Self-supervised cad reconstruction by learning sketch-extrude operations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2023; pp. 16816-16826.

67.

Chang

Sung

Goel

Lambourne

Birdal

, et al. Point2cyl: Reverse engineering 3d objects from point clouds to extrusion cylinders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; 2022; pp. 11850-11860.

68.

Kociecki

Adeli

. Two-phase genetic algorithm for topology optimization of free-form steel space-frame roof structures with complex curvatures. Engineering Applications of Artificial Intelligence.2014; 32: 218-227.

69.

Fan

Yan

. 3D reconstruction based on hierarchical reinforcement learning with transferability. Integrated Computer-Aided Engineering.2023; (Preprint): 1--13.

70.

Fan

Song

. Look inside 3D point cloud deep neural network by patch-wise saliency map. Integrated Computer-Aided Engineering.2024; (Preprint): 1-16.

71.

Zheng

Yang

Leng

Cheng

Xie

, et al. A survey of feature modeling methods: Historical evolution and new development. Robotics and Computer-Integrated Manufacturing.2020; 61: 101851.

72.

Breitenberger

Apostolatos

Philipp

Wüchner

Bletzinger

. Analysis in computer aided design: Nonlinear isogeometric B-Rep analysis of shell structures. Computer Methods in Applied Mechanics and Engineering.2015; 284: 401-457.

73.

Andrade

Trevelyan

Leonel

. A NURBS-discontinuous and enriched isogeometric boundary element formulation for two-dimensional fatigue crack growth. Engineering Analysis with Boundary Elements.2022; 134: 259-281.

74.

Amos

Yarats

. The differentiable cross-entropy method. In: International Conference on Machine Learning. PMLR; 2020; pp. 291-302.

75.

Arjovsky

Chintala

Bottou

. Wasserstein generative adversarial networks. In: International conference on machine learning. PMLR; 2017; pp. 214-223.

76.

Xue

Tong

Neri

Zhang

. PEGANs: Phased Evolutionary Generative Adversarial Networks with Self-Attention Module. Mathematics.2022; 10(15): 2792.

77.

Tang

Liu

Duan

. MATR: Multimodal Medical Image Fusion via Multiscale Adaptive Transformer. IEEE Transactions on Image Processing.2022; 31: 5134--5149.

78.

Schwan

Schenck

. A three-step model for the detection of stable grasp points with machine learning. Integrated Computer-Aided Engineering.2021; 28(4): 349-367.

79.

Sørensen

Nielsen

Karstoft

. Routing in congested baggage handling systems using deep reinforcement learning. Integrated Computer-Aided Engineering.2020; 27(2): 139-152.

80.

Xue

Zhu

Neri

. A self-adaptive multi-objective feature selection approach for classification problems. Integrated Computer-Aided Engineering.2022; 29(1): 3-21.

81.

Nogay

Adeli

. Detection of epileptic seizure using pretrained deep convolutional neural network and transfer learning. European neurology.2021; 83(6): 602-614.

82.

Nogay

Adeli

. Diagnostic of autism spectrum disorder based on structural brain MRI images using, grid search optimization, and convolutional neural networks. Biomedical Signal Processing and Control.2023; 79: 104234.

83.

Nogay

Adeli

. Multiple Classification of Brain MRI Autism Spectrum Disorder by Age and Gender Using Deep Learning. Journal of Medical Systems.2024; 48(1): 15.

84.

Zhang

Xue

Neri

Cai

Slowik

. Multi-Objective Self-Adaptive Particle Swarm Optimization for Large-Scale Feature Selection in Classification. International Journal of Neural Systems.2024; 2450014-2450014.

85.

Chen

Deng

Furukawa

. Polydiffuse: Polygonal shape reconstruction via guided set diffusion models. Advances in Neural Information Processing Systems.2024; 36.

86.

Eddahab

FZA

Horváth

. Recommendation function for smart data analytics toolbox to support semantic merging of middle-of-life data streams. In: 2020 International Conference on Control, Automation and Diagnosis (ICCAD); 2020, pp. 1-7.

87.

Zhao

Shen

. A federated distillation domain generalization framework for machinery fault diagnosis with data privacy. Engineering Applications of Artificial Intelligence.2024; 130: 107765.

88.

Kociecki

Adeli

. Shape optimization of free-form steel space-frame roof structures with complex geometries using evolutionary computing. Engineering Applications of Artificial Intelligence.2015; 38: 168-182.

89.

Aldwaik

Adeli

. Advances in optimization of highrise building structures. Structural and Multidisciplinary Optimization.2014; 50: 899-919.

A parametric and feature-based CAD dataset to support human-computer interaction for advanced 3D shape learning

Abstract

Keywords

1. Introduction

2.1 B-rep-based CAD dataset

Table 1 Differences of features among WHUCAD, DeepCAD and Fusion 360 Gallery

3.4 Selection feature to learn and imitate HCI process under deep learning network

(1) How the selection and name mechanism work when extruding a sketch into a 3D model

(2) How the selection and name mechanism work when conducting a chamfer operation on an edge of the box.

(3) How the selection and name mechanism work when a face is split by a previous operation.

3.6 Discussion of feature types, history length and type distribution

6.2 Industrial experiments

Footnotes

Acknowledgments

References

Table 1
Differences of features among WHUCAD, DeepCAD and Fusion 360 Gallery