Qixing Huang: 2014 Networks of Shapes, Images, and Programs: The Power of Joint Data Analysis Workshop

 

Wednesday, April 16, 2014
Location: Fisher Conference Center, Arrillaga Alumni Center

"Labeling Large Shape Collections"
11:30am - 12:00pm

Abstract:

Thanks to innovations in 3D scanning and modeling technologies, large collections of 3D shapes become widely available. For example, the popular Google/Trimble 3D Warehouse repository contains millions of man-made 3D shapes, and the size continues to grow at a high rate. To make these big geometric datasets useful, we typically start by classifying them into different classes. So far most existing works have focused on classifying shapes into different high-level categories, e.g., cars, chairs, desks, etc. However, for large shape collections, the shapes within each category still exhibit significant variation. For example, chair models from the 3D Warehouse contain dozens of sub-classes, including chairs-with-arms, swivel chairs, rocking chairs, etc. These fine-grained classes contain rich semantic information that can benefit a variety of applications such as product search, browsing and exploration of shape variability, and interactive shape modeling.


In this talk we consider the problem of classifying a collection of shapes of the same category into fine-grained classes. We start from understanding the challenges of fine-grained shape classification, including (i) lacking canonical domains to compare shapes, (ii) subtle geometric features that characterize each class, and (iii) the fact that only sparse and noisy labels are available. We then discuss a semi-supervised classification pipeline which combines joint shape alignment in common spaces, joint class-specific distance metric learning and graph-based semi-supervised classification. The performance of this pipeline turns to be much better than existing shape classification techniques. We will also discuss various applications of fine-grained shape classification.


Bio:

Qixing Huang is a postdoctoral researcher at the Computer Science department of Stanford University. His main research interests are in organizing large collections of geometric data and exploring various data-driven applications in analysis, animation, modeling and visualization. Qixing Huang received the PhD degree from Stanford University in 2012, working with Prof. Leonidas Guibas. He is the recipient of Mr. and Mrs. Chin-Nan Chen Stanford Graduate Fellowship from 2008-2011, and the recipient of the Best Paper Award at the conference Symposium on Geometry Processing 2013. He has served on program committees of SGP, PG, SPM and GMP.