Faculty Recruiting Support CICS

Representation Learning for Shape Decomposition, By Shape Decomposition

10 May
Monday, 05/10/2021 8:30am to 10:30am
Zoom Meeting
PhD Dissertation Proposal Defense
Speaker: Gopal Sharma

Zoom link: https://umass-amherst.zoom.us/j/92715272065

Abstract: The ability to parse 3D objects into their constituent parts is essential for humans to understand and interact with the surrounding world. Imparting this skill in machines is important for various computer graphics, computer vision, and robotics tasks. Machines endowed with this skill can better interact with its surrounding, perform shape editing, texturing, recomposing, tracking, and animation. In this thesis, we ask two questions. First, how can machines decompose 3D shapes into their fundamental parts? Second, does the ability to decompose the 3D shape into these parts help learn useful 3D shape representations?

We start by exploring different ways in which a shape can be decomposed into compact representations. We are interested in approximating the shape with compact representations that are used by the expert in graphics modeling packages like MAYA, Inkscape, Illustrator, etc., and can be easily used by non-expert users to manipulate the shape. In this thesis, we focus on parsing the shape into constructive solid geometry programs and parametric surface patches using deep neural networks.

Advances in network architecture to process 3D data such as point clouds, meshes and voxels have pushed the boundaries of 3D recognition and generation tasks. Training these models requires a large amount of labeled datasets. We explore ways to alleviate this problem by relying on shape decomposition methods to guide the learning process. We first study the use of freely available decomposition, albeit inconsistent, from shape repositories to learn 3D shape features. Later, we show that learning to decompose a 3D shape into primitives also helps in learning shape representation. We claim that decomposing shape into compact representation is a fundamental task that has applications in shape editing and shape recognition tasks. We further show improvement over state-of-the-art approaches in these areas.

Co-Chairs, Subhransu Maji and Evangelos Kalogerakis
Member, Rui Wang
Outside Member, Siddhartha Chaudhuri