Chapter 11: Collaborative Product Development 2: Digital Design

Title: Validation of Purdue engineering shape benchmark clusters by Crowdsourcing

Author(s): Ananda Prasanna Jagadeesan, Jan Wenzel, Jonathan Corney, Xiu Yan, Andrew Sherlock, Carmen Torres-Sanchez, William Regli

Address: Department of Design, Manufacture and Engineering, Management (DMEM), University of Strathclyde, James Weir Building, 75 Montrose St, Glasgow, Gl 1XJ, UK | Department of Mechanical Engineering, University of Edinburgh, Kings Building, Edinburgh, EH9 3JY, UK | Department of Design, Manufacture and Engineering, Management (DMEM), University of Strathclyde, James Weir Building, 75 Montrose St., Glasgow, Gl 1XJ, UK | Department of Design, Manufacture and Engineering, Management (DMEM), University of Strathclyde, James Weir Building, 75 Montrose St., Glasgow, Gl 1XJ, UK | ShapeSpace Ltd, 26 George Square, Edinburgh, EH8 9LD, Scotland | Department of Mechanical Engineering, School of EPS, Heriot-Watt University, Edinburgh, EH14 4AS, UK | Drexel University, Department of Computer Science, 3201 Arch Street, Philadelphia, PA 19104, USA

Reference: International Conference on Product Lifecycle Management 2009 pp. 549 - 562

Abstract/Summary: The effective organization of CAD data archives is central to PLM and consequently content based retrieval of 2D drawings and 3D models is often seen as a 'holy grail' for the industry. Given this context, it is not surprising that the vision of a 'Google for shape', which enables engineers to search databases of 3D models for components similar in shape to a query part, has motivated numerous researchers to investigate algorithms for computing geometric similarity. Measuring the effectiveness of the many approaches proposed has in turn lead to the creation of benchmark datasets against which researchers can compare the performance of their search engines. However to be useful the datasets used to measure the effectiveness of 3D retrieval algorithms must not only define a collection of models, but also provide a canonical specification of their relative similarity. Because the objective of shape retrieval algorithms is (typically) to retrieve groups of objects that humans perceive as 'similar' these benchmark similarity relationships have (by definition) to be manually determined through inspection. This paper reports the methodology developed to employ a commercial 'Crowdsourcing' service to distribute the task of assessing the similarity of models in a 3D dataset to anonymous workers on the internet. To determine the effectiveness of this distributed approach it was applied to the class of 107 'flat-thin wall components' within the 'Purdue engineering shape' benchmark's collection of 3D models. The resulting families of similar shapes (defined at varying resolutions) identified within the dataset, show close correspondence with Purdue's published clusters. This validates the use of Crowdsourcing as fast, cheap and effective method of content classification for CAD data.

Order a copy of this article