Automatic CAD Assemblies Generation by Linkage Graph Overlay for Machine Learning Applications

Abstract

Enlarging 3D model databases by shape synthesis is a large field of research. Indeed, the use of machine learning techniques requires a huge amount of labeled CAD models, and it is therefore crucial to rely on large and varied databases. Most of existing works in shape synthesis focus on everyday life objects generation [1]. However, these methods often do not work on assemblies composed of several CAD models, and it is the aim of this paper to develop a new shape synthesis method to enlarge existing CAD assembly databases. Today, there exist lots of free databases of non-labeled CAD models (e.g. GrabCAD, 3D Warehouse, Turbosquid) often available as STEP or IGES files. Unfortunately, very few of these databases are labeled. Other databases like PartNet [2] and ShapeNet [3] are currently labeled by crowdsourcing, but they do not contain complex mechanical assemblies. Current works in shape synthesis often use auto-encoders to generate new coherent CAD assemblies [4]. Moreover, there are probabilistic models to create diversity in large 3D Database [5] or to classify 3D assemblies [6]. Those techniques often use linkage graphs [7] to classify and generate new coherent assemblies. Furthermore, information within the linkage graphs differs according to the method, and those graphs are not suitable for complex CAD assemblies like hydraulic pumps. But the main issue of all methods is still that those databases have to be labelled. The method explained in this paper consists in creating new labeled CAD assemblies from existing ones by linkage graph overlay. Here, the STEP file format has been adopted in order to be the most reproductible and to be adaptable. The linkage graphs are automatically created thanks to the identification of the linkages between the components. Indeed, linkages are not included in the STEP files and they need to be computed. Theses linkage graphs are then analyzed and components with similar linkages are detected. Finally, once the similarities detected, the corresponding components can be exchanged to created new assemblies for which the labels can be directly inherited from the source assemblies. The contribution is threefold: (i) a method to create linkage graphs from existing non-labelled CAD assemblies; (ii) a method to recognize basic components using linkage graphs; (iii) a smart overlay method to replace some components while keeping the coherence between all the components of the assembly. The algorithm has been implemented in Python on FreeCAD and it has been tested on several test cases. Figure 1 shows the overview of the method, from the graph synthesis to the components overlay, finishing with the replacement of the components. The results are presented and discussed, and a conclusion ends this extended abstract while discussing the next steps.