3D Data Mining: Part and Information Re-Use in a PLM context: Research Paper Introduction (RPI) - Par : Roland Maranzana,

3D Data Mining: Part and Information Re-Use in a PLM context: Research Paper Introduction (RPI)

Roland Maranzana
Roland Maranzana Profil de l'auteur(e)
Laurent Maranzana est professeur au Département de la production automatisée à l’ÉTS. Ses recherches portent sur la gestion du cycle de vie des produits, la fabrication et la conception assistée par ordinateur, la métrologie et l’usinage.


A research paper introduction is a blog article presenting a research paper done by researchers from École de technologie supérieure (ÉTS) de Montréal.


Fig. 1. Example of 3D drawings of miscellaneous parts that could be reuse. Source [Img1]


Fig. 2. Various type of 3D CAD drawings and files. Source [Img1]

To remain competitive, companies favor re-using their data (fig. 1). In the aerospace or automotive sectors, products are very complex and are the result of interactions from various departments (design, process planning, manufacturing, etc.).


The participants (partners, contractors, etc.) often belong to different companies in various locations. The enormous quantity of information generated during multiple product development projects must be stored in a vault (or electronic database) where its permanence and accessibility can be assured. Accessibility is not enough, however, to allow the reuse of information. It is also necessary to be able to quickly locate those parts which are potentially relevant to the new project and to be able to assess their relevance with precision [1]. And data today are on 3D! Engineering and manufacturing information is related to geometric data (fig. 2).

But what about searching and finding the right CAD file? How to find 3D models?

Many methods exist to do search and find function. The most known method is text based search. Unfortunately, this method is inefficient to find CAD files. Describing parts using words is not easy: in some case the same description is given to different parts, in other, similar parts have differnet descriptions (fig. 3). Text based search is language dependant and requires knowledge of metadata values. And for most of all, text fields are kept blank!


Fig. 3. An example of a text based search for a bracket. Source [Img1]

Text based, alpha-numerical and semantic search rely on metadata and have the same drawbacks. Specification tree mining would be interesting but require that all data have the specification tree done.

The best approach is to use 3D drawings to find 3D drawings! The fundamental idea behind 3D search consists of findind CAD files similar to a reference 3D model based on their shape and size. Research made for 3D search focuses mainly on descriptors, storage of descriptors, query interfaces, descriptors comparison and results display (fig. 4).


Fig. 4. Process to index CAD files. Source [Img1]

3D search has been a very active area of research since 2000 cause by the increasing need for 3D search engines in different domains:

  • Web 3D content search
  • Engineering
  • Molecular biology
  • Medecine
  • Paleontology
  • Forensics
  • Gaming
  • Etc.


Descriptors are abstraction and formalization of a 3D model. They should have the following properties:

  • Discriminating (sensitive, unique)
  • Concise to store
  • Quick to compute
  • Efficient to match
  • Robust and stable to model degeneracies
  • Invariant to translations and rotations

However, a descriptors with all these properties is hard to develop. Several types of descriptors are proposed in scientific papers:

  • Spherical harmonics: this technique decomposes a 3D model into a collection of functions defined on concentric spheres.  It uses spherical harmonics to discard orientation information (phase) for each one (fig. 5). It makes a shape descriptor orientation invariant and descriptive [2];

Fig. 5: Spherical harmonics descriptor process. Source [Img2]

  • D2Shape distributors: with this technique, the shape of a 3D model is represented as a probability distribution sampled from a shape function measuring geometric properties of a 3D model (fig. 6). One such shape distribution called D2 represents the global shape of an object by the probability distribution of Euclidean distances between pairs of randomly selected points on its surface [2];

Fig. 6: D2 distributions for 5 tanks (gray curves) and 6 cars (black curves). Source [Img2]

  • Topology based approach (graph based and skeleton): the general idea is to derive 1D skeletal curve from a 3D object such that each curve represents a significant part of the object (fig. 7).  These curves are then converted to an attributed graph representation (a skeletal graph), which can be used for indexing, matching, segmentation, correspondence finding, etc.;

Fig. 7. A plane model, its medial axis, the voxelized centerline representation, and the resulting skeletal graph. Source [Img2]

  • Invariant global parameters: for [3], those invariant descriptors constructed using the methods of invariant theory for 3D object recognition “are shape descriptors that are unaffected by object pose, by perspective projection and by the intrinsic parameters of the camera.”
  • 2D visuals descriptors (content based image retrieval methods): those descriptors are descriptions of the visual features of the contents (shape, color, texture, etc.). The shape of a 3D object can be approximately represented with a limited number of 2D shapes taken from different viewing directions (fig. 8).


Fig. 8: 2D shapes taken from a chair in 3D. Source [Img1]

Descriptors comparison

Different comparison methods ara proposed depending of the descriptor used. Computing a distance between descriptors is the most common method used to retrieve 3D models, classify a 3D model and find the closest neighbors (filter out).

Query interfaces and results display

The query interfaces and display results depends of the descriptor used. It could be a:

  • 2D sketching (vector or image) (fig. 9);

Fig. 9. 2D shapes painted with a pixel paint program with set of matching objects returned by the system. Source [Img2]

  • 3D sketching (graphic or CAD type) or an existing 3D model with iterative refinement (fig. 10);

Fig.10: A query made for a a cup sketched and its five closest matches. Source [Img2]

  • Parts browser search result (fig.11);

Fig. 11. 3D parts search result example. Source [Img1]

  • Geolus search result: A shape search application that allows manufacturers to quickly locate 3D models of digitally defined parts from large heterogeneous data sources on the basis of geometric similarity [4] (fig. 12).


Fig. 12. Geolus search results example for a part. Source [Img1]

RM1logoApproach developed for 3D search

Following a research done with CRIAQ, we have developed a 3D based search engine focusing on CAD files, mechanical parts and engineering activities. The search engine had to be fast, robust, reliable and usable.

The system developed is named 3DPart Finder and is commercially available.  Regarding technical specifications, the search engine is shape and size sensitive. It distinguishes mirror parts and identifies duplicates with high confidence and accuracy (fig. 13). It is integrated to CAD and PLM systems and adapted to large number of CAD files (> 1M files).


Fig. 13. Mirror part and identical part search. Source [Img1]

The descriptor designed is directly extracted from native 3D model (Catia V5, NX, Inventor, Pro/E, SolidWorks, Solid Edge). It takes advantage of accessing B-rep information, has N dimension vector corresponding to morphological and dimensional features. For descriptor comparison, different methods are used to define a distance between descriptors: duplicate and mirror parts identification and it consider mirror parts as identical and as distinct (fig. 14).


Fig. 14. Duplicate and mirror parts identification. Source [Img1]

The query interface uses CAD systems to define the reference 3D model to start searches (fig. 15). Queries are made with an existing or a rough 3D model. The results are displayed on the CAD systems since it’s integrated.


Fig. 15. The query interface. Source [Img1]

Search use CATIA functions to analyze similar parts and make final selection. Search results are presented as an assembly (fig.16).


Figure 16. Search results presentation. Source [Img1]

This research was presented at the Global Product Interoperability Summit 2013 on September 9-12, at the Sheraton Wild Horse Pass Resort in Chandler, AZ, USA.

RmaranzanaDr. Roland Maranzana is a professor at École de technologie supérieure of Montreal (ÉTS), Quebec, Canada and the CTO of 3DSemantix company.


[toggle title= »References »]

[1] Msaaf O., R. Maranzana and L. Rivest (2007). Part Data Mining for Information Re-use in a PLM Context. Proceedings of GT2007 ASME Turbo Expo 2007: Power for land, Sea and Air, May 14-17, 2007, Montreal, Canada.

[2] Dobkin and al. (2013). Princeton Shape Retrieval and Analysis Group. Princeton University, NJ, USA. Retrieved 11/21/2013, World Wide Web.

[3] Forsyth D. and al. (1991). Invariant Descriptors for 3-D Object Recognition and Pose. IEEE Transactions on pattern and Machine Intelligence, vol. 13, no 10, pp. 971-991.

[4] Rowe J. (2007). Geolus Search, The Google of 3D. Retrieved 11/21/2013, World Wide Web.





Roland Maranzana

Profil de l'auteur(e)

Laurent Maranzana est professeur au Département de la production automatisée à l’ÉTS. Ses recherches portent sur la gestion du cycle de vie des produits, la fabrication et la conception assistée par ordinateur, la métrologie et l’usinage.

Programme : Génie de la production automatisée 

Laboratoires de recherche : LIPPS – Laboratoire d'ingénierie des produits, procédés et systèmes 

Profil de l'auteur(e)


    Laisser un commentaire

    Votre adresse de messagerie ne sera pas publiée. Les champs obligatoires sont indiqués avec *