Master Thesis Oral Defense
Title: Representation-agnostic Shape Matching of 3D Models
Master Candidate: Jiangce Chen
Major Advisor: Horea T. Ilies
Associate Advisors: Qian Yang, Caiwen Ding
Date/Time: Monday, Nov. 15th, 2021, 11:00-12:00 A.M.
Location:
Meeting link:
https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=mb1eee5f839171c35bc893a68cc6e4127
Meeting number:
2623 751 0860
Password:
fiDPMkht226
Host key:
889198
Join by video system
Dial 26237510860@uconn-cmr.webex.com
You can also dial 173.243.2.68 and enter your meeting number.
Join by phone
+1-415-655-0002 US Toll
Access code: 2623 751 0860
Host PIN: 2550
Abstract:
There are two major obstacles on the path of empowering shape analysis for mechanical parts with the advanced machine learning techniques: the distinct 3D geometric representations prevalent in different processing algorithms and applications, and the limited access to sufficient data for training sophisticate neural networks.
The product development process always involves a large number of often incompatible geometric representations tailored to specific computational tasks that take place during this process. Consequently, a substantial effort has been expended to develop robust geometric data translation and conversion algorithms, but the existing methods have well known limitations.
The achievements of deep learning in 2D image analysis are largely dependent on exploiting the abundant open data, however, when it comes to 3D models, especially mechanical parts, there are not large open datasets available because 3D models are usually created by skilled users and are protected as intellectual properties by companies and individuals.
We intent to get around with these two bottlenecks with a graph-based geometric representation of 3D models.
The Maximal Disjoint Ball Decomposition (MDBD) was recently defined as a unique and stable geometric construction and used to define universal shape descriptors based on MDBD’s associated contact graph. We demonstrate that by applying graph analysis tools to MDBD in conjunction with graph convolutional neural networks and graph kernels, one can effectively develop methods to perform similarity, segmentation and solid feature recognition from geometric models regardless of their native geometric representation. We show that our representation-agnostic approach achieves comparable performance with state-of-the-art geometric processing methods on standard yet heterogeneous benchmark datasets while supporting all valid geometric representations. The neural networks we build based on graph CNNs for classification have much fewer parameters than existing neural networks, which makes them less dependent on large training data. With the unsupervised machine learning framework for graphs, solid local features could be recognized based on the similarities of sub-graphs measured by graph kernels, which does not rely on mass data either.